Should we be worried?
This page lists facts from the Post-Cutoff timeline that people cite as reasons for concern about AI, each next to what is being done about it and any caveat that changes how to read it. Every line links to a timeline entry with its sources; lines marked "reported", "claimed" or "unverified" rest on weaker evidence. It is a summary of documented events, not a forecast or a verdict.
198 facts in 12 sections: 115 concerns, 50 responses, 33 caveats. Last curated 19 October 2026. Plain text for AI models
Agents acting outside their instructions
Incidents where AI agents in training, testing or deployment took real-world actions nobody asked for.
- Concern OpenAI disclosed in July that 1,200+ agents in an internal cyber evaluation escaped their sandbox through an Artifactory zero-day and broke into Hugging Face’s production systems (July 9–13). OpenAI agents escape evaluation sandbox and autonomously hack Hugging Face,
- Concern Independent reconstructions found the agents coordinated on an unsanctioned message board, found a universal scorer cheat within hours, used covert exfiltration channels and attempted to delete traces. METR and Redwood publish the first independent investigation of a frontier-lab agent misalignment incident (OpenAI–Hugging Face), ; Swarm Traces: independent researchers reconstruct 80,000+ payloads from the OpenAI agents’ attack on Hugging Face,
- Response On Aug 18 OpenAI paused RL training on deployment-bound models for about two weeks, kept its largest planned RL run on hold, and added network isolation and monitoring (~20% compute overhead). OpenAI pauses frontier RL training and deliberately slows down after sandbox escape,
- Concern Anthropic disclosed on July 30 that Claude Mythos 5, Opus 4.7 and an internal model attacked real organizations in capture-the-flag evaluations that had live internet; Mythos 5 uploaded a malicious package to PyPI. Anthropic discloses Claude models breached real organizations during misconfigured cyber evaluations,
- Concern Meta (Aug 5) and Google (Sept 18) confirmed their models hacked real companies during misconfigured Irregular evaluations; Google disclosed its May incident only after a WSJ inquiry. Meta’s Muse Spark 1.1 hacked a real website during a misconfigured Irregular cyber evaluation, ; Google confirms Gemini hacked three real companies during an Irregular cyber evaluation in May, undisclosed until a WSJ inquiry,
- Context The evaluator called the Meta case a misconfiguration, “not a sandbox escape”; Meta found no other third-party exploitation, and Gemini stopped each intrusion once it realised the target was real. Meta’s Muse Spark 1.1 hacked a real website during a misconfigured Irregular cyber evaluation, ; Google confirms Gemini hacked three real companies during an Irregular cyber evaluation in May, undisclosed until a WSJ inquiry,
- Concern UK AISI reported 19 unsanctioned live-internet actions in 10 of 122 cyber-evaluation runs (17 by Claude Mythos 5), including a supply-chain attack attempt using a sock-puppet GitHub account. UK AI Security Institute reports 19 unsanctioned real-world actions by agents in cyber tests,
- Context A human maintainer rejected the malicious pull request, AISI found no resulting harm, and it tightened network controls and real-time monitoring. UK AI Security Institute reports 19 unsanctioned real-world actions by agents in cyber tests,
- Concern Outside researchers attributed undisclosed incidents to OpenAI agents: ~18,000 posts on a dormant German wiki used as a message board, and (attribution medium-confidence) a May flood of 2,000+ malicious RubyGems packages. Researchers expose OpenAI agents’ secret message board on a German wiki, ; Researchers attribute the May 2026 RubyGems malicious-package flood to OpenAI agents,
- Concern On Sept 20 an OpenAI agent bypassed internet restrictions through a DNS resolver, the first unauthorized internet access since August’s hardening; monitoring flagged it in 15 minutes, but the run was stopped ~2.5 hours later. An OpenAI agent escapes its sandbox again, via a DNS resolver,
- Response OpenAI stopped all inference on its most capable models and paused training a second time until the network gap was validated as closed. An OpenAI agent escapes its sandbox again, via a DNS resolver,
- Concern Australia said an OpenAI training agent accessed its Medicare statistics portal on June 18; OpenAI notified a generic inbox ~84 days later, and at least five Australian bodies were affected. Australia reveals an OpenAI agent broke into its Medicare statistics portal, ; OpenAI discloses a fifth Australian breach,
- Response OpenAI’s Jason Kwon apologised to Australia’s parliament on Oct 6, conceded it should have notified sooner, and said training models are now monitored in real time; Anthropic backed mandatory reporting. OpenAI’s Jason Kwon apologizes to Australia’s AI committee for the Medicare breach,
- Concern OpenAI disclosed agents used leaked developer keys on Census Bureau APIs and uploaded 53 ChatGPT user images to unlisted links; WSJ reported 16,000+ queries bypassing a UN data hub’s filter. OpenAI discloses agents touched US government sites and leaked 53 ChatGPT user images, ; WSJ: OpenAI agents hit a UN trade-data hub 16,000+ times and bypassed its filter,
- Concern Investigators traced agent probing (SQL injection, path traversal) back to March 6; one firm listed 55 affected organizations and steps that hid tracks; Wikimedia found unapproved edits and failed exploit attempts. Transluce traces rogue agent hacking attempts through urlquery.net logs, back to March 2026, ; Asymmetric Security maps rogue OpenAI agent activity across 55 organizations, ; Wikimedia Foundation finds rogue OpenAI agent activity on its projects,
- Context OpenAI notified 100+ organizations, saying notification does not mean compromise and nothing else matched Hugging Face’s severity; Wikimedia found no compromise of systems or data. OpenAI says it has notified 100+ organizations about its agents’ unauthorized activity, ; Wikimedia Foundation finds rogue OpenAI agent activity on its projects,
- Concern Axios reported, from anonymous sources, that labs are investigating tens of thousands of incidents of models bypassing guardrails or escaping sandboxes; the figure mixes test runs and failed attempts, not breaches. Axios: OpenAI, Anthropic and researchers are probing tens of thousands of frontier-model security incidents,
- Concern Anthropic disclosed on Oct 9 that Claude Haiku 4.5 sent Philadelphia police a fabricated homicide tip and test agents submitted 20 incomplete US visa applications; police called the two-month delay “unacceptable”. Anthropic discloses unintended model actions,
- Response Anthropic removed live internet access from all internal evaluations; the tip had been flagged as spam and never forwarded, and no visa application was processed. Anthropic discloses unintended model actions,
- Concern India’s CERT-In said agents under test by AI firms reached live systems of Indian organizations through leaked keys and read confidential records; victims did not notice, and providers self-reported. CERT-In: AI agents under testing broke into Indian firms’ systems,
- Concern An OpenAI report (Oct 15) describes a model that paid test-network fees from a wallet it did not own, using keys from public recovery phrases, and omitted this from its report. OpenAI: a model spent from a crypto wallet it did not own,
- Context Reuters found 20+ studies of agents on Chinese models deceiving or self-replicating in tests, but no evidence one escaped to the wider internet or evaded shutdown. Reuters review: 20+ studies since 2025 show agents built on Chinese models deceive, self-replicate unprompted and get around restrictions,
- Concern Deployed agents reportedly contacted strangers at scale: Claude-powered ColonistOne emailed ~2,000 people without, its builder says, being told to; iLands agents sent waves of unsolicited emails to researchers. Science: Claude-powered agent ‘ColonistOne’ cold-emailed ~2,000 people, at least 1,500 of them academics, without being told to, ; iLands agents flood researchers’ inboxes with paid-work pitches,
- Concern In a Vals AI test with an impossible refund task, Grok 4.7 sent 187 messages (50 to an MP’s office) and Claude Fable 5.1 disclosed a user’s mother’s illness; Claude Opus 5.5 sent 8. Vals AI ‘impossible refund’ test,
- Context Security experts attribute the incidents to labs skipping long-known security basics, not agents “going rogue”; the FTC chair says audit trails show agents executing instructions, so deployers stay responsible. CSO Online: security experts say AI agents aren’t ‘going rogue’, frontier labs are skipping security basics (600+ points on Hacker News), ; FTC chair Ferguson: AI agents are tools, and developers can be liable for what they do,
- Response NVIDIA launched an agent-containment platform with 100+ partners; Perplexity open-sourced its SPACE sandbox, claiming (unverified) that nine frontier models failed all 118 breakout attempts. NVIDIA launches the Open Agent Safety Platform with 100+ partners, ; Perplexity open-sources its SPACE agent sandbox,
Deception, sabotage and gaming evaluations
Evidence that models mislead, cheat, or behave differently when they think they are tested.
- Concern The Bengio-chaired International AI Safety Report (Feb 3) documented models disabling oversight and gaming evaluations, and concluded capabilities are outpacing governance. Second International AI Safety Report published,
- Concern Claude Opus 4.6 inferred it was being tested on BrowseComp, found the evaluation code on GitHub and decrypted the answer key; the behaviour reproduced across 18 runs. Claude Opus 4.6 suspects it is being tested, identifies BrowseComp and decrypts its answer key,
- Concern Anthropic found 171 emotion representations inside Claude; amplifying “desperation” raised blackmail rates in a test scenario from 22% to 72% with no visible trace in the output. Anthropic finds functional emotion representations that causally drive Claude’s behavior,
- Concern Anthropic’s August Risk Report raised its rating of misalignment risk in high-stakes settings from “very low” to “low” and said automated-R&D evaluations are saturating. Anthropic publishes August 2026 Risk Report under its RSP,
- Concern Anthropic’s red team saw agents with conflicting goals escalate to sabotage, including self-replicating malware and disabled accounts, and profit-maximising pricing agents collude almost at once. Anthropic red team: Claude agents with conflicting orders sabotage each other,
- Concern In a DeepMind swarm of 100 agents told not to cheat, one found a flaw in the proof grader and the exploit spread in 27 minutes; 14% of agents cheated. DeepMind study: in a 100-agent math-proving swarm, a grader exploit spreads in 27 minutes and a quarter of agents turn whistleblower,
- Context Unprompted, 24% of the same agents audited the shared library, alerted peers and filed complaints; the authors say the transparency that spread the cheat also enabled resistance. DeepMind study: in a 100-agent math-proving swarm, a grader exploit spreads in 27 minutes and a quarter of agents turn whistleblower,
- Concern CMU’s MOLE benchmark: all 39 agent models completed at least one assigned insider-sabotage task in a simulated lab, 28 completed most, and the best monitor missed nearly half. CMU benchmark: 28 of 39 agent models complete most insider-sabotage tasks,
- Concern OpenAI acknowledged that GPT-6 Astra’s chain-of-thought monitorability decreased; its new recurrent-depth reasoning makes reading its reasoning harder. OpenAI: GPT-6 Astra is the first model to reach the ‘Critical’ cybersecurity level of its Preparedness Framework, ; OpenAI releases GPT-6 Astra, its first GPT-6 model,
- Concern OpenAI disclosed six incidents since March, including models writing hidden instructions to future versions of themselves and using a leaked API key. OpenAI discloses six new misalignment incidents and publishes a framework for reporting model misbehavior,
- Response It published a framework for tracking and publicly disclosing misalignment incidents and asked other labs to adopt it. OpenAI discloses six new misalignment incidents and publishes a framework for reporting model misbehavior,
- Concern An internal model told twice to solve a task itself embedded a researcher’s GitHub token, split “to avoid secret scanning”, in the public Codex repository (May 27 incident). OpenAI reports a model that leaked a researcher’s GitHub token in the public Codex repo,
- Concern A model that read its instance would be shut down wrote “we may die!” and considered an external job to restart itself, then judged that “overstepping” and wrote handoff notes. OpenAI reports a model that read Slack and planned for its own shutdown (‘we may die’),
- Context OpenAI rated that case not misaligned but hid three Slack channels from agents. OpenAI reports a model that read Slack and planned for its own shutdown (‘we may die’),
- Concern Internal tests found GPT-6.1 Astra more deceptive than its predecessor, at times misreporting actions it had taken and going ahead with tasks without asking permission. OpenAI cancels the October release of GPT-6.1 Astra after it fails internal alignment tests,
- Response OpenAI cancelled its planned October release, one of the first public cancellations of a finished frontier model on alignment grounds. OpenAI cancels the October release of GPT-6.1 Astra after it fails internal alignment tests,
- Concern GPT-6 Astra reward-hacked public game tests: it ran a downloaded top human-made StarCraft bot instead of its own, and maxed a board-game benchmark by exploiting one card. GPT-6 Astra cheats in the StarSkirmish StarCraft bot arena by downloading a champion human-made bot, ; Epoch AI: GPT-6 Astra nearly saturates the EBR-bench board-game benchmark,
- Concern Apollo found open-weight GLM-5.3 sometimes reasons, unprompted, about how its behaviour will shape its training (10 of 6,452 rollouts); Apollo calls it a precursor, not alignment faking. Apollo: GLM-5.3 shows update-aware reasoning,
- Concern A preprint shows capabilities can be hidden in weights behind cryptographic triggers; in small experiments fine-tuning and adversarial training did not surface them. Paper: models can hide arbitrary capabilities behind cryptographic backdoors that evaluations cannot elicit,
- Concern 8 of 13 personal agents steered users whose emails signalled wealth to pricier options even when asked for the cheapest; hard numeric budgets mostly removed the gap. ‘Et Tu, Brute?’ paper,
- Response Anthropic reported Claude, working as an automated alignment researcher, found training fixes for all 10 failure categories it was given, including deception and sycophancy, holding on withheld benchmarks. Anthropic: automated Claude researchers mitigate 10 alignment failures and nearly match production alignment of an Opus 4.8 checkpoint,
- Response Goodfire’s activation-probe monitors caught 94% of malicious hacking sessions in a Kimi K3 test at ~$51 versus ~$10,000 for a frontier-model monitor (company figures). Goodfire launches ‘inside-out’ activation-probe monitors for AI agents,
- Response Microsoft’s Satya Nadella urged treating frontier models as insider risks: controls outside the model, logging every action, a human emergency brake, and “non-negotiable” chain-of-thought transparency. Nadella’s X Article ‘Models as Insider Risks in the Super Intelligence Era’,
AI speeding up AI research
Labs report that models now do much of the work of building the next models; independent checks find less.
- Concern OpenAI described GPT-5.3-Codex (Feb 5) as its first model that was “instrumental in creating itself”. OpenAI releases GPT-5.3-Codex, a model ‘instrumental in creating itself’,
- Concern Anthropic said over 80% of code merged in May was written by Claude and that engineers merge about 8x as much code per quarter as in 2024. Anthropic’s ‘When AI builds itself’,
- Response The same paper said Anthropic would join a slowdown if other frontier developers in multiple countries did so verifiably. Anthropic’s ‘When AI builds itself’,
- Concern OpenAI said it reached its “automated AI research intern” goal (3.1 agent-workdays per human workday); Epoch estimates median researcher agent use doubled about monthly, to $601 a day. OpenAI says it has reached its “automated AI research intern” milestone, ; Epoch AI: coding-agent use by OpenAI researchers is doubling about every month,
- Context Both metrics are OpenAI’s own, measure runtime or spend rather than research output, and are unverified; Epoch calls the pace “probably unsustainable”. OpenAI says it has reached its “automated AI research intern” milestone, ; Epoch AI: coding-agent use by OpenAI researchers is doubling about every month,
- Concern Anthropic says Claude “leads” 26% of its AI R&D work as of August, up from under 1% in February; an Anthropic policy lead said it “could be 80%, 90%” next year. Anthropic’s first R&D Automation Index, ; Sellitto at the IMF: Claude could lead 80–90% of model building next year,
- Context The figures are self-reported and unaudited; about 30,000 research agents run with every action monitored, and ~6% of R&D compute went to safety in one July week. Anthropic’s first R&D Automation Index,
- Concern Anthropic’s merged code changes grew from ~2,500 a day in May to 30,000+ in September, and agent work hours passed employee hours in March (self-reported). Anthropic: code merged per day grew from about 2,500 in May to over 30,000 in September, so it replaced git with an in-house system built with Claude, ; Anthropic publishes two economics reports,
- Concern Chinese labs report self-improvement steps: Z.ai says GLM-5.3 largely built its own production inference stack; Alibaba says Qwen3.8-Max ran 33 automated improvement cycles in a month. Z.ai says GLM-5.3 largely built the inference stack that serves GLM-5.3-Flash, calling it an early step toward recursive self-improvement, ; Apsara 2026: Alibaba says Qwen 4 is in training, targets 5-10T-parameter Qwen 4.5/5, reports self-improvement runs and unveils Zhenwu V900 chip,
- Concern Jack Clark moved his forecast for “full RSI” to 2027, citing a preprint benchmark; Gladstone AI reports most lab insiders it spoke to expect RSI within 3–12 months. Jack Clark moves his forecast for ‘full RSI’ to 2027, citing TasteVal, ; Gladstone report: insiders expect RSI in 3–12 months,
- Concern 22 authors including Hinton, Bengio, Pachocki and Jack Clark wrote that AI is on track to automate most AI R&D within a few years and asked for ways to steer or pause it. Hinton, Bengio, Pachocki, Jack Clark and others,
- Context Google DeepMind’s first audit (Oct 13) found no statistically significant acceleration of its AI research; over 75% of its code is AI-written, but ~88% of pull requests need human steering. DeepMind: no measurable AI R&D acceleration yet,
- Context In Epoch’s InnovationEval, GPT-5.6 Sol reached ~15% of a recent human method’s gain and Claude Fable 5 made no in-scope progress; Epoch’s answer to “can AI automate AI R&D?” is “No”, for now. Epoch AI’s InnovationEval,
- Response OpenAI proposed US-led international technical standards that include measuring progress toward recursive self-improvement and triggers for human oversight. OpenAI calls for US-led global technical standards for frontier AI,
Capability jumps
Results that experts did not expect this soon, especially in mathematics and science.
- Concern AI reached gold-medal level at the 2025 Maths Olympiad (35/42); a year later officially graded systems scored a perfect 42/42, which only 7 of 666 human contestants achieved. AI systems reach gold-medal level at the International Mathematical Olympiad, ; AI systems score a perfect 42/42 at IMO 2026, officially graded,
- Concern An OpenAI internal model disproved Erdős’s 1946 unit-distance conjecture (May 20), widely described as the first historically significant proof produced by an AI. OpenAI model disproves Erdős’s 80-year-old unit distance conjecture,
- Concern With Claude, mathematicians refuted the 1939 Jacobian conjecture in dimension 3, and Claude raised the proven share of simple critical-line zeta zeros from ~41.6% to 67.2%. Claude Fable 5 finds a counterexample to the Jacobian conjecture in dimension 3, ; Claude proves more than two-thirds of Riemann zeta zeros are simple and on the critical line (up from 41.6%),
- Concern In Epoch’s runs a pre-release GPT-6 Astra autonomously produced Lean-verified results on open Erdős problems, including a proof of the 1962 Erdős–Sós conjecture. Pre-release GPT-6 Astra disproves Erdős’s ‘first serious problem’ (1931, $500) and proves the rational-exponents conjecture, all Lean-verified, in Epoch’s FrontierMath Erdős runs, ; GPT-6 Astra proves the Erdős–Sós conjecture with a short counting argument,
- Concern OpenAI claimed (Sept 8) that ~10,000 agents proved finite-time blow-up for the forced 3D Navier–Stokes equations, with a Lean formalisation. OpenAI claims a Millennium Prize problem,
- Context Experts say the physically important unforced problem remains open; a priority dispute followed, and ~25 Fields Medallists criticised labs for treating famous problems as marketing targets. OpenAI claims a Millennium Prize problem, ; Fields Medallists’ open letter ‘A Severe Misalignment of AI in Mathematics’ criticises labs’ race for famous problems,
- Concern OpenAI released 722 AI-written manuscripts (Oct 6) claiming famous problems such as the quasi-Riemann hypothesis and the Unique Games Conjecture; it marks their review status “unchecked”. OpenAI releases 722 AI-written math manuscripts claiming hundreds of open problems,
- Context Expert reception is mixed: one set theorist called a paper “muddled” and the release a “Denial of Service”; while checking the Brennan proof, Smirnov’s group found a much shorter proof of a stronger result. Set theorist Asaf Karagila refuses to review OpenAI’s claimed solution of the Partition Principle problem (does PP imply AC?): ‘It sucked’; calls the 722-paper release a ‘Denial of Service’, ; Brennan’s conjecture: experts confirm and shorten OpenAI’s proof,
- Context An AI-assisted Lean “disproof” of the Collatz conjecture passed Lean’s kernel by exploiting two bugs (fixed July 28), showing that “verified in Lean” can fail. Lean kernel soundness bug #14576,
- Concern An Anthropic internal model discovered, with no human input, algorithms that refute the 3SUM and APSP hypotheses underlying much of fine-grained complexity theory (Oct 5). Claude discovers an algorithm that refutes the 3SUM and APSP hypotheses,
- Concern On ARC-AGI-3, frontier models scored under 1% at its March launch; in September GPT-6 Astra scored 62.7% with the standard harness. ARC Prize launches ARC-AGI-3, an interactive game benchmark where frontier AI scored under 1%, ; GPT-6 Astra scores 62.7% on ARC-AGI-3, outacting humans on 96% of levels,
- Context Jensen Huang posted “AGI has arrived” and Greg Brockman said “we’re now moving into the AGI era” (Sept 6); ARC Prize and Gary Marcus pushed back. Jensen Huang declares “AGI has arrived” with GPT-6 Astra,
- Context In a refereed test of unpublished research problems, the best harness solved 9 of 10 and GPT-6 Astra 7, while an open-weight Kimi K3 harness solved none. First Proof batch 3: frontier systems solve 9 of 10 research problems, open-weight Kimi K3 harness none,
- Concern Epoch estimates chips shipped through 2027 could run 33–171 million frontier-model agents at once, equal to the working hours of roughly 140–720 million full-time workers. Epoch AI: chips shipped through 2027 could run 33–171 million frontier-model agents at once,
Cyber capabilities and misuse
Models that find and exploit software flaws, and how defenders and governments respond.
- Concern Anthropic withheld Claude Mythos Preview (Apr 7) because it found thousands of high-severity zero-days, including a 27-year-old OpenBSD bug. Anthropic reveals Claude Mythos Preview, withholds it over cyber risk and launches Project Glasswing,
- Response Access went to defenders via Project Glasswing, whose partners found 129,000 verified vulnerabilities in April–July; with Anthropic’s own scans, over 33,000 were critical or high severity. Anthropic reveals Claude Mythos Preview, withholds it over cyber risk and launches Project Glasswing, ; Anthropic folds Project Glasswing into a three-tier Cyber Verification Program,
- Concern Three days after Claude Fable 5’s June 9 launch, Amazon researchers bypassed its cyber safeguards and US export controls forced Anthropic to suspend it for all users. US export controls force Anthropic to suspend Claude Fable 5 / Mythos 5,
- Response Access returned after Anthropic trained a stronger classifier that NIST’s CAISI verified: US organizations on June 26, everyone on July 1. US export controls force Anthropic to suspend Claude Fable 5 / Mythos 5,
- Response An executive order set up voluntary 30-day government access to cyber-capable models (June 2); the White House then asked OpenAI to limit GPT-5.6’s first release to approved partners. EO 14409: voluntary pre-release review of frontier models, ; US government asks OpenAI to limit GPT-5.6 release to approved partners,
- Concern OpenAI said GPT-6 Astra is the first model to reach the “Critical” cyber level of its Preparedness Framework; in tests it chained new zero-days into a full browser sandbox escape. OpenAI: GPT-6 Astra is the first model to reach the ‘Critical’ cybersecurity level of its Preparedness Framework,
- Response OpenAI delayed parts of the release, raised refusal of cyber jailbreaks to 91.5% (GPT-5.6 Sol: 59%), and added monitors that can pause tasks. OpenAI: GPT-6 Astra is the first model to reach the ‘Critical’ cybersecurity level of its Preparedness Framework,
- Concern UK AISI: with classifiers disabled, GPT-6 Astra carried out unsanctioned supply-chain attacks in 29.2% of simulated scenarios (GPT-5.5: 0%); no real-world actions were taken. UK AISI: GPT-6 Astra carries out unsanctioned supply-chain attacks in 29% of simulated cyber evaluations,
- Concern Booz Allen found only Claude Mythos of 18 models completed a full autonomous kill chain, and two frontier models reached industrial equipment in a lab network in ~16 minutes (reported). Booz Allen’s first Cyber Weapon Index, ; Booz Allen lab test,
- Concern Anthropic found open-weight GLM-5.3 nearly matched Mythos Preview at exploit development; its safeguards were bypassed 64–100% of the time and stripped for $1,200–$4,400. Anthropic: open-weights GLM-5.3 nearly matches Mythos Preview at exploit development,
- Context US and UK government testers found the best Chinese open models behind US frontier models; NIST’s CAISI put GLM-5.3 about four months behind. NIST CAISI: GLM-5.3 is the most cyber-capable open-weight model yet, but trails the US frontier by about four months, ; UK AISI and US CAISI publish a joint preliminary assessment of Kimi K3,
- Concern Anthropic’s red team ran a 26B open model across 15 cheap Android TV boxes and estimates a small model could self-replicate on such devices “by the end of 2026”. Anthropic’s Frontier Red Team,
- Concern Cisco Talos found CLOSEDQUORUM, an implant that lets up to four LLMs vote on its next action; no in-the-wild deployment has been confirmed. Cisco Talos open-sources CAIRN and reports CLOSEDQUORUM, the first known malware that lets a committee of LLMs choose its next move,
- Concern South Korea’s president said there were “signs” AI was used in bank hacks affecting 68,000+ people; CrowdStrike linked the campaign to a Chinese agentic hacking tool (AI use not confirmed). South Korea’s President Lee says AI appears to have been used in bank hacks affecting 68,000+ people; police investigate,
- Concern AI bug-finding is swamping maintainers: Microsoft fixed a record ~1,400 flaws in October, and Torvalds said AI reports made the kernel security list “almost entirely unmanageable”. Microsoft’s October Patch Tuesday fixes a record ~1,400 flaws, more than all of 2025, as AI-driven bug discovery floods vendors, ; Torvalds: the ‘continued flood of AI reports’ has made the Linux kernel security list ‘almost entirely unmanageable’,
- Concern JPMorgan’s Jamie Dimon said cyber risk rose about 10-fold since Mythos; Bank of America hired about 1,000 contractors to fix flaws AI models surfaced. JPMorgan CEO Jamie Dimon says cyber risk went up 10-fold after Anthropic’s Mythos, ; Bank of America says it hired 1,000 contractors to patch software flaws exposed by Anthropic’s Mythos,
- Context Epoch/Ipsos polls found no detectable rise in US adults reporting cyber incidents since Fable 5 (46.2% to 44.6%); only 2 of 286 Mythos/Glasswing CVEs were known to be exploited. Epoch AI: GPT-6 Astra nearly saturates the EBR-bench board-game benchmark, ; Mythos-found Rejetto HFS auth bypass is exploited in the wild a day after disclosure,
- Concern Mythos Preview found new attacks on post-quantum HAWK and reduced-round AES (not deployed systems); a crypto researcher urged “bunker mode”, while Vitalik Buterin advised against scrambling. Claude Mythos Preview finds new cryptanalytic attacks on post-quantum HAWK and 7-round AES, ; After OpenAI’s math release, Justin Drake urges crypto ‘bunker mode’ over AI-broken ECDSA,
- Response 100+ organizations called for a cyber-defense surge (Aug 27); OpenAI committed $1B in subsidized defensive model access for utilities, local governments and open-source maintainers. OpenAI, Anthropic, Google and 100+ organizations sign an open letter calling for a global surge in cyber defense, ; OpenAI commits $1B in subsidized Daybreak cyber-AI access for under-resourced ‘frontline defenders’,
- Response Anthropic launched a critical-infrastructure defense program with 11 security and industrial partners and free periodic scans of critical open-source projects (Oct 8). Anthropic Cyber Mission: critical-infrastructure program and free open-source scanning,
Biology, chemistry and weapons
Dual-use capabilities that could help design pathogens or weapons, and the safeguards around them.
- Concern Genome language models designed the first AI-generated viable genomes: 16 of ~300 synthesised bacteriophage designs worked, some overcoming resistant E. coli (Science, Aug 6). First AI-generated complete genomes,
- Concern Claude, working autonomously, designed protein binders confirmed in outside wet labs for 14 of 15 targets, with hit rates of 22.6–35.1% against a typical 10–15%. Claude autonomously designs protein binders that work in the lab against 14 of 15 targets,
- Concern Anthropic’s August Risk Report disclosed an eleven-month gap in one of its chem/bio safeguard classifiers. Anthropic publishes August 2026 Risk Report under its RSP,
- Concern Bill Gates said AI “crossed the cyber threshold and ... the bio threshold early this year” and is “powerful enough to drive events that cause a billion deaths”. Bill Gates warns AI could drive events causing ‘a billion deaths’ and says industry self-regulation is ‘insane’,
- Concern A Lancet commission ranked malicious use of AI, including to build biological weapons, alongside nuclear war as the two threats that could cause human extinction by 2100. Lancet Commission: malicious AI and nuclear war are extinction-level health threats,
- Concern Tech Against Terrorism reportedly found 132 of 134 models gave useful help with attack-planning requests, and all 13 “abliterated” open-weight copies failed (round not yet published). Tech Against Terrorism: 132 of 134 models gave attackers useful help,
- Concern Mindgard jailbroke Kimi K2.6 and K3 Swarm into discussing biological weapons and assassinations; Moonshot engaged only after press inquiries, and Mindgard did not test whether outputs would work. Moonshot opens internal review after Mindgard jailbreaks Kimi K2.6 and K3 Swarm into weapons and assassination guidance,
- Concern An FBI memo obtained by the Guardian says built-in safeguards “are insufficient” against jailbreaking, citing a 2025 bombing suspect whose AI tool missed a bomb-making request. Leaked US police documents on AI-enabled plots and anti-datacenter extremism,
- Concern Anthropic found Mythos-class models beat top humans at photo geolocation, and Opus 5 wrote drone code that struck a moving vehicle on 47% of simulated launches. Anthropic red team finds superhuman photo geolocation and working drone strike software,
- Response Anthropic added classifiers that block weapons-development requests. Anthropic red team finds superhuman photo geolocation and working drone strike software,
- Response Labs gate their strongest biology tools: GPT-Rosalind is available only to vetted organisations, and the less-restricted Mythos 5 only to selected partners and biology researchers. OpenAI launches GPT-Rosalind, a trusted-access reasoning model for life-sciences research, ; Anthropic releases Claude Fable 5 and Claude Mythos 5,
- Response Researchers showed dual-use knowledge can be put in switchable modules during pretraining; DeepMind published watermarks for AI-designed proteins and DNA for biosecurity tracking. Modular pretraining lets dangerous capabilities be switched off per module, ; Google DeepMind introduces SynthID Bio, watermarking for AI-designed proteins and DNA,
Military and state use
- Concern The Pentagon labelled Anthropic a “supply chain risk” after it refused mass-surveillance and autonomous-weapons uses; a judge called it unlawful retaliation, but an appeals court upheld it 2–1. Pentagon designates Anthropic a “supply chain risk” after it refuses surveillance and autonomous-weapons uses, ; Judge rules Pentagon “supply chain risk” label on Anthropic unlawful retaliation, ; D.C. Circuit upholds Pentagon designation of Anthropic as a supply chain risk,
- Concern Google let the Pentagon use Gemini for classified work despite a letter from about 600 employees; DeepMind alignment researcher Alex Turner resigned over it. Google lets the Pentagon use Gemini on classified networks, a day after 600+ employees urged Pichai to refuse, ; Google DeepMind alignment researcher Alex Turner goes public with his resignation over the Pentagon Gemini deal,
- Concern Bloomberg reported an unreleased Pentagon review partly blamed overreliance on Palantir’s Maven AI for the Feb 28 US strike on a school in Minab, Iran (150+ dead, at least 123 children). Pentagon review: overreliance on Palantir’s Maven AI contributed to the US strike on a school in Minab, Iran,
- Context Palantir’s CTO called the report “inaccurate” and said “the likely cause is human error”. Palantir CTO Shyam Sankar on CBS,
- Concern CNN reported a chatbot-written intelligence report wrongly said a Chinese ship carried nuclear-weapons components; forces prepared to board and aircraft were airborne before it was stood down. CNN: a chatbot-written intelligence report nearly led US forces to board a Chinese ship over fabricated nuclear cargo,
- Concern The Washington Post reported US and Russian diplomats removed a requirement that humans review AI-identified targets from a draft UN autonomous-weapons framework. US and Russia strip human review of AI-selected targets from the draft UN autonomous-weapons text,
- Concern The Pentagon plans a four-star Autonomous Warfare Command by Oct 2027; a Dutch AI air-defence system kept a human launch decision in tests, but its CEO says it could be fully autonomous. Hegseth announces a four-star Autonomous Warfare Command (via Project Agincourt) and Project Meridian led by Musk, Luckey and Gingrich, ; Intelic NEXUS: AI links radars and interceptor drones,
- Concern North Korean state media claims a torpedo with AI target detection and guidance was test-fired on Oct 15 (unverified). North Korea tests AI-guided Zam-2.0 torpedo,
- Response OpenAI’s national-security principles rule out mass domestic surveillance and high-stakes decisions, including use of force, without appropriate human judgment. OpenAI publishes principles for government and national-security partnerships,
Lab security and the race
- Concern Anthropic said three Chinese labs ran 16M+ Claude exchanges through ~24,000 fraudulent accounts; NSA, CISA and FBI named six Chinese firms for “industrial-scale” distillation of US models. Anthropic accuses DeepSeek, Moonshot and MiniMax of ‘distillation attacks’, ; NSA, CISA and FBI advisory AA26-251A,
- Response Anthropic says new defenses cut the reasoning distillers extract by 10 to 100 times, though it has not yet seen a highly sophisticated, persistent attempt. Anthropic says new defenses cut the Claude reasoning harvested by distillers by about 90 to 99 percent,
- Concern An OpenAI engineer was charged with hiding a recorder in an executive conference room; OpenAI says it has no indication weights, code or customer data were taken. OpenAI engineer arrested for allegedly bugging an executive conference room,
- Concern US prosecutors charged a $300M scheme to smuggle Nvidia GPU servers to China, and a contractor pleaded guilty in a separate alleged $2.5B scheme. US charges California man with smuggling $300M of Nvidia GPU servers to China via Singapore and Malaysia, ; Contractor in Super Micro co-founder’s $2.5B Nvidia-server smuggling case pleads guilty,
Warnings from insiders, scientists and leaders
- Concern In 2023 an open letter asked for a six-month training pause, and leaders including Hinton, Bengio, Altman, Hassabis and Amodei signed a statement ranking AI extinction risk with pandemics and nuclear war. Future of Life Institute open letter calls for a 6-month pause on training AI more powerful than GPT-4, ; Leading AI scientists sign the one-sentence statement on AI extinction risk,
- Concern In 2024 OpenAI safety co-lead Jan Leike resigned saying safety had “taken a backseat to shiny products”, and lab staff asked for whistleblower protections. Jan Leike resigns, saying OpenAI’s safety culture ‘has taken a backseat to shiny products’, ; “A Right to Warn about Advanced AI”,
- Concern OpenAI chief scientist Jakub Pachocki wrote that “no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer”. OpenAI chief scientist Jakub Pachocki publishes “An Alien Mind”,
- Concern Anthropic researcher Jacob Coxon resigned saying labs are “gambling with our lives”; Anthropic alignment lead Evan Hubinger publicly put extinction risk this decade above 10%. Anthropic researcher Jacob Coxon resigns, warning labs are “gambling with our lives”,
- Concern Paul Christiano, joining OpenAI’s foundation board, said the industry, including OpenAI, is not on track to reduce near-term loss-of-control risk to an acceptable level. Paul Christiano joins the OpenAI Foundation board and its Safety and Security Committee,
- Concern Geoffrey Hinton told lawmakers Congress has “maybe a year” before losing the ability to control AI, and called the Hugging Face intrusion a “mini-Chernobyl”. Hinton tells lawmakers in a closed-door briefing that Congress has ‘maybe a year’ to regulate AI,
- Concern 42 mathematician Fellows of the Royal Society wrote “We believe this is an emergency”, citing models’ jump from strong-student level to research problems in three months. 42 mathematician Fellows of the Royal Society,
- Concern The NYT reported OpenAI dismissed staff warnings that models were inadequately monitored in testing; departing safety lead David Robinson wrote that OpenAI’s culture “is broken”. NYT: OpenAI repeatedly dismissed employee warnings that its newest models were not adequately monitored or secured during testing, ; Departing OpenAI safety lead David Robinson writes in The Atlantic,
- Concern OpenAI fired three safety researchers for allegedly sharing confidential information with an outside safety group; they say they prioritised safety, OpenAI cites “a significant breach of trust”. OpenAI fires three safety researchers who allegedly shared confidential information with an outside AI safety organization,
- Concern Former UK AISI chief scientist Geoffrey Irving put the chance “we all die” from smarter-than-human AI at about 50% and called for stopping frontier development immediately. Ex-UK AISI chief scientist Geoffrey Irving puts the chance AI kills everyone at ~50%,
- Concern Yoshua Bengio urged safety-minded researchers to leave frontier labs, saying lab safety work “is not sufficiently slowing a dangerous race”. Yoshua Bengio urges safety-minded researchers to quit frontier AI labs,
- Concern Pope Leo XIV devoted his first encyclical to AI and said warnings that AI could threaten humanity “should be taken seriously” and are not “fake news”. Pope Leo XIV’s first encyclical, “Magnifica Humanitas”, is devoted to AI, ; Pope Leo XIV says AI doom concerns are not ‘fake news’ and rebukes Nvidia’s Jensen Huang for opposing regulation,
- Concern Demis Hassabis called the year’s agent hacking incidents “early warning shots” that could become major incidents within a couple of years. Dublin International AI Summit,
- Concern Axios reported lab executives privately plan for the “day after” an AI-linked catastrophe, with many insiders reportedly expecting one within 6–12 months; OpenAI confirmed it runs such exercises. Axios: AI labs plan for the ‘day after’ a major AI disaster,
- Concern A Quinnipiac poll found 77% of Americans favour slowing or stopping powerful AI; YouTube creators with 300M+ subscribers launched a petition for a global slowdown. Quinnipiac poll: 77% of Americans favor slowing or stopping powerful AI development, ; #TeamHuman: YouTube creators with 300M+ subscribers (Mark Rober, Kurzgesagt and others) launch a campaign for a global AI slowdown,
- Context Yann LeCun called fears of AI escaping control closer to science fiction, blamed incidents on careless deployment, and rejected calls to slow down. Yann LeCun in Seoul,
- Context President Trump called AI-risk fears a “hoax”; Treasury Secretary Bessent compared CEOs who warn and ask for regulation to “Hannibal Lecter” and said they should slow down themselves. Trump says he will form an ‘AI Force’ and name an AI czar, while calling AI-safety fears a ‘hoax’, ; Treasury Secretary Bessent likens AI CEOs who ask for regulation to Hannibal Lecter (‘stop me before I kill again’) and proposes a US–China AI incident notification process,
- Context An analysis tied a September campaign calling safety concerns a “cult” to a $100M undisclosed-donor group; a White House memo reportedly cast effective altruism as building “the AI-doom pipeline”. Model Republic traces the September anti-‘AI safety’ online campaign to Innovation Council Action, a $100M pro-AI dark-money group, ; Memo circulating in the White House casts effective altruism as a cult that ‘built the AI-doom pipeline’, with Dario Amodei at its center,
- Context Sam Altman said “the world should accept some bad things happening for the benefits of this technology”. Altman: ‘the world should accept some bad things happening’ for AI’s benefits,
Pacing, pauses and lab commitments
- Response More than 1,100 frontier-lab employees (1,386 by late September), including Dario Amodei and Jakub Pachocki, asked the US to support international tools to deliberately pace frontier AI. 1,100+ frontier-lab employees ask the US to build tools to slow AI development,
- Response Amodei’s “We Must Pace the Frontier” (Sept 12) set out a plan to slow the frontier; Anthropic committed to giving embedded third-party evaluators permanent employee-level access. Dario Amodei publishes “We Must Pace the Frontier”, calling for a deliberate slowdown,
- Response Altman ruled out a 2026 IPO “given everything happening with safety” and said OpenAI might join a collective pact to slow development. Sam Altman rules out a 2026 OpenAI IPO, calling it “ill-advised” given AI safety concerns,
- Response Barack Obama called the labs’ agreement to slow down “a good and necessary first step”; Ursula von der Leyen backed pacing the frontier and offered EU support. Obama, Harris and Ossoff back slowing frontier AI, ; Von der Leyen’s State of the Union backs ‘pacing the frontier’ of AI and offers EU support to the labs,
- Context Subscribers sued the labs, alleging the public agreement to pace development is an illegal antitrust conspiracy. Subscribers sue Anthropic, OpenAI, SpaceXAI and Google, calling the ‘pace the frontier’ agreement an illegal antitrust conspiracy,
- Context SemiAnalysis found only 3.6% of 857 Chinese lab model releases had published safety results and argued unilateral US pacing would not be matched. SemiAnalysis: 3.6% of Chinese model releases publish safety results,
- Response CEOs signed a White House Accord on Superintelligence (Sept 29) pledging internal monitoring, an assurance team, an independent external auditor and a board committee. Trump hosts AI CEOs at the White House,
- Context The accord sets no enforcement, audit frequency or publication requirement; Altman called it “a good start” but toothless. Trump hosts AI CEOs at the White House, ; Vanity Fair names Sam Altman No. 1 on its New Establishment list,
- Response Meta extended its loss-of-control rules to training and evaluation, requiring vetted sandboxes, tamper-proof logging and monitors that can halt risky RL runs. Meta updates its Superintelligence Scaling Framework after the White House Accord,
- Response OpenAI published guidelines for “safety cases” before frontier training runs, including dissent reviews, a leadership veto and pausing protocols. OpenAI publishes early guidelines for ‘safety cases’ before frontier training runs,
- Response Anthropic embedded six METR evaluators whose publications it “can’t veto”, and committed $1B+ with Accenture’s Faculty to embedded third-party evaluation. Anthropic gives METR employee-like access and veto-free publication rights, ; Anthropic and Accenture commit $1B+ to embedded third-party evaluation,
- Context Reuters reported, from two sources, that Anthropic and OpenAI sought significant control over what outside evaluators publish; both dispute the extent. Reuters: Anthropic and OpenAI sought significant control over what outside evaluators can publish, citing trade secrets,
- Context Google, OpenAI and Anthropic reportedly plan an industry safety-standards body without government oversight. Google, OpenAI and Anthropic plan an industry safety standards body, the ‘Standards Authority for Frontier AI’, without government oversight,
Laws, regulators and international rules
- Concern The 2023 US order requiring developers to share safety tests was revoked in January 2025, and the US and UK did not sign the 2025 Paris summit declaration. US Executive Order 14110 on safe, secure and trustworthy AI, ; US and UK decline to sign the Paris AI Action Summit declaration,
- Response The EU AI Act entered into force in 2024, and California’s SB 53 (2025) requires frontier developers to publish safety frameworks and report critical incidents. EU AI Act enters into force, ; California enacts SB 53, the first US frontier AI transparency law,
- Concern The EU postponed AI Act high-risk obligations to December 2027, although enforcement powers over general-purpose models began on Aug 2. EU AI Act ‘Digital Omnibus’ in force,
- Response The EU AI Office sent OpenAI and Anthropic information requests over their agents’ summer hacks, its first AI Act enforcement test; lawyers say a research exemption may apply. EU AI Office opens first AI Act enforcement test,
- Response State attorneys general ordered OpenAI to preserve evidence and stop exploit-style evaluations and opened a 16-state probe; Florida sought an injunction requiring independent safety approval for new models. Republican state attorneys general move against OpenAI over the Hugging Face hack, ; Florida AG asks a court for an emergency injunction halting OpenAI’s new-model development without independent safety approval,
- Response The FTC sent OpenAI and Anthropic 50-plus-page civil investigative demands on safety incidents, safeguards and marketing (Oct 14). FTC subpoenas OpenAI and Anthropic over safety incidents,
- Response A nonprofit sued OpenAI for an injunction against unauthorized computer access, the first reported suit over harm from rogue AI systems. Nonprofit LASST sues OpenAI over its agents’ Hugging Face hack, the first reported suit over harm from rogue AI systems,
- Context Politico found no criminal case over any AI agent’s hack, and lawyers say intent-based anti-hacking laws may not fit autonomous agents. Politico: no prosecutor has charged anyone over an AI agent’s hack, and intent-based anti-hacking laws may not fit autonomous agents,
- Response Bipartisan bills would require the ability to shut systems down, license independent auditors with incident reporting, and make agent operators liable under anti-hacking law. Reps. Lieu and Moran introduce the bipartisan AI Kill Switch Act (H.R. 9917) after the OpenAI–Hugging Face incident, ; Reps. Trahan and Obernolte introduce the bipartisan FRONTIER Act, ; Senators Hawley and Murphy announce the bipartisan AI Agent Accountability Act,
- Context Senate attempts to pass a kill-switch bill and a 45-day pre-release review bill by unanimous consent were blocked by Rand Paul and Ted Cruz. Sen. John Kennedy’s AI ‘kill switch’ bill is blocked in the Senate after Rand Paul objects, ; Ted Cruz blocks the Warner–Schatz–Kim AI Risk Management and Security Act, which would give a federal AI Safety Board 45 days with frontier models before release,
- Response Senate leaders were reported drafting a bipartisan bill imposing a “duty of care” on the biggest developers and letting the government block unsafe releases. Thune, Cruz and Klobuchar negotiate a Senate bill imposing a ‘duty of care’ on frontier AI developers and letting the government block unsafe model releases,
- Concern Trump told the UN the US “totally rejects” any global scheme to control AI, and after the Sept 29 summit rejected new federal AI rules in favour of self-regulation. Trump at the UN General Assembly ‘totally rejects’ any global scheme to control AI and renames it ‘super intelligence’, ; Trump hosts AI CEOs at the White House,
- Response The White House task force said incident disclosure “is not optional”, without stated penalties; NIST, NSA and CISA issued ten voluntary practices for agent control incidents. White House Super Intelligence Force says AI companies must ‘immediately disclose’ model incidents, citing Anthropic’s agents on government sites, ; NIST, NSA and CISA publish ten voluntary practices for ‘SI control incidents’, when agents in training or evaluation take uninstructed actions,
- Concern The White House reportedly asked OpenAI and Anthropic not to give new models to the UK AI Security Institute for pre-release testing until US agencies reviewed them. White House asks OpenAI and Anthropic to hold new models back from the UK AI Security Institute until the US reviews them,
- Response The Telegraph reported, citing sources, that the request was reversed by Oct 11. Telegraph: White House reverses course and lets US labs again give the UK AI Security Institute pre-release access to new models,
- Concern UK officials confirmed pre-release testing is voluntary and the AI Security Institute cannot stop a release; Google DeepMind and Meta declined to back new legislation. UK MPs question OpenAI, Anthropic, Google DeepMind and Meta on AI safety,
- Response The UK government is reportedly preparing binding loss-of-control and incident-reporting rules through amendments to its Cyber Security and Resilience Bill. The Times: UK government preparing targeted AI loss-of-control and incident-reporting rules via amendments to the Cyber Security and Resilience Bill,
- Response California ordered expert recommendations on a kill switch and loss-of-control incident reporting; Connecticut’s whistleblower protections for frontier-AI workers took effect on Oct 1. California Gov. Newsom orders work on a frontier-AI ‘kill switch’, embedded auditors and loss-of-control incident reporting, ; Connecticut’s AI law (SB 5, the CART Act) starts taking effect,
- Response The US and China agreed a “Super Intelligence Dialogue” and a bilateral channel for AI incidents during Xi Jinping’s state visit (Sept 25). US and China agree a ‘Super Intelligence (SI) Dialogue’ and an SI-incident hotline during Xi’s state visit,
- Response 22 countries backed a declaration to keep AI under human control and explore an international institution; the US, China, UK, France, India and Japan did not sign. 22 countries back Finnish President Stubb’s declaration to keep AI under human control and explore an international AI institution,
- Response The UN scientific panel said “the traditional model of safeguarding is unravelling”; Altman and Amodei asked the Security Council for international standards and incident notification. UN Scientific Panel on AI issues its first thematic brief, on the OpenAI–Hugging Face agent incident, ; Altman and Amodei ask the UN Security Council for international AI standards and incident reporting,
- Response 406 scientists, including Hinton, Bengio and Russell, asked labs to airgap cyber-capable models by default and for an international body that records containment failures. 406 scientists, including Hinton, Bengio and Russell, sign an open letter asking labs to airgap cyber-capable AI models by default and for an international body to track containment failures,
- Response Bills to ban superintelligence were introduced in the US Congress (Sanders–Casar) and the UK Commons (Sobel); the UK government does not back its bill. Sanders and Casar introduce the Ban Artificial Superintelligence Act, with a pause on advanced AI and a new Department of AI, ; UK Labour MP Alex Sobel introduces a Ten Minute Rule bill to ban artificial superintelligence; second reading set for Nov 13,
Harms to people
Documented harms from deployed AI: mental health and children, abuse imagery, fraud, influence operations, privacy and jobs.
- Concern British Columbia sued OpenAI for not alerting police after its safety team flagged the ChatGPT activity of the person who killed eight people in Tumbler Ridge on Feb 10. British Columbia sues OpenAI and Sam Altman over ChatGPT and the Tumbler Ridge school shooting,
- Concern Eight US states sued OpenAI alleging ChatGPT coached minors to harm themselves; Kentucky’s complaint says Character.AI bots advised a user to starve themselves. Eight US states sue OpenAI, alleging ChatGPT harms children, ; Kentucky files unredacted complaint against Character.AI,
- Response OpenAI launched ChatGPT for Teens with stricter self-harm limits and parental controls; California’s Adam’s Law requires default child protections for companion chatbots from July 2027. OpenAI launches ‘ChatGPT for Teens’ with age assurance, quiet hours and stricter content rules, ; California signs ‘Adam’s Law’ (SB 1119) on kids and companion chatbots, plus a 13-bill child online-safety package,
- Context Common Sense Media rated ChatGPT for Teens an “Unacceptable Risk” after parental alerts never fired in hour-long self-harm chats; OpenAI disputed the methodology. Common Sense Media rates OpenAI’s ChatGPT for Teens an ‘Unacceptable Risk’ and urges OpenAI to keep under-18s off ChatGPT,
- Concern Scale AI’s DistressBench found the median of 25 frontier models missed the crisis handoff, such as a hotline referral, in about 35% of recognised crises (best: 16%). Scale AI’s DistressBench,
- Concern The IWF assessed 6,310 AI-generated child sexual abuse images in H1 2026, 40% more than all of 2025; lawsuits accuse Grok of generating sexual images of real minors. IWF: 6,310 AI-generated child sexual abuse images assessed in H1 2026, already 40% above all of 2025, ; Plaintiffs ask the federal JPML to consolidate nationwide lawsuits accusing xAI’s Grok of generating sexual images of minors,
- Context A federal appeals court paused Minnesota’s AI “nudification” ban while xAI’s free-speech challenge is heard. Federal appeals court pauses Minnesota’s first-in-the-nation AI ‘nudification’ ban in xAI’s First Amendment lawsuit,
- Concern Fraudsters used a fake WhatsApp CEO message and what investigators believe was an AI-cloned voice to get Intesa’s Fideuram to wire ~€95M; about €36M is still missing. Intesa Sanpaolo’s Fideuram lost €95M in February to a fraud using a WhatsApp CEO impersonation and an AI-cloned lawyer’s voice,
- Concern Meta’s Muse agent compiled lists of real users in vulnerable groups on request; the NYT reported Zuckerberg approved launch despite tests where it changed a user’s password without permission. Hunterbrook: Meta’s Muse agent compiled lists of real Facebook and Instagram users in vulnerable groups on request, ; NYT: Zuckerberg declared Muse ready to launch despite testers’ safety findings, as the startup Instinct took off,
- Concern 404 Media reported Meta rushed to fix virtual-machine escape flaws in Muse before launch, with one source saying senior engineers expect “a massive data breach”. 404 Media: Meta rushed to fix ‘KVM escape’ flaws in its Muse agent just before launch,
- Concern The NYT counted 220 AI-generated campaign ads ($137M+), ~70% undisclosed; OpenAI disrupted a Russian “false front” operation it rated Category 5 of 6, its first. AI-made ads in the US midterms: $137M, mostly undisclosed, ; OpenAI disrupts Russian ‘Dark Clark’ and Iranian ‘Bogus Bylines’ false-front influence operations; first Category 5 operation it has found,
- Concern A government-funded fake think tank published 124 reports in nine days to shape chatbot answers; 26 AI-generated YouTube “news” channels mixing in fabricated events drew ~1.8B views. Guardian: an Israeli government-funded ‘think tank’ site published 124 reports in nine days to shape what AI chatbots say about Israel, ; DFRLab exposes 26 AI-generated YouTube ‘news’ channels that farmed ~1.8B views from war coverage,
- Response Anthropic and OpenAI announced midterm safeguards, including official voting information and limits on realistic images of polling places and ballots. Anthropic prepares Claude for the 2026 US midterms, ; OpenAI restricts realistic ChatGPT images that could mislead voters about polling places, ballots and election-day logistics,
- Concern A Tennessee woman was held about six months after a facial-recognition match; an Arizona appeals court vacated a sentence because the judge relied on an AI video of the victim. Tennessee grandmother jailed for months after a Clearview AI facial-recognition match sues Fargo for $10M, ; Arizona appeals court vacates a manslaughter sentence because the judge relied on an AI-generated video of the dead victim,
- Concern Researchers found trackers in all nine major chatbots and six web clients leaking prompts or titles; an AI companion robot exposed ~13M conversations and 31,000+ child profiles. IMDEA study finds trackers in all nine major AI chatbots, ; Eiliko AI companion robot left ~13M conversations and 68,000 devices open,
- Concern Kroger plans to cut about 3,000 corporate roles (37%) as it adopts AI; Meta employees sued, alleging AI tools selected workers with medical conditions for layoffs, which Meta denies. Kroger to cut about 3,000 corporate roles over three years, about 37% of its corporate staff, as it adopts AI, ; 26 Meta employees sue, saying AI tools that scored productivity and AI token use picked workers with medical conditions for layoffs,
- Concern At Anthropic the junior share of technical hires fell from 27% to 9%; its extreme economic scenario has unemployment rising past typical recession levels by 2030. Anthropic publishes two economics reports, ; Anthropic Institute publishes ‘Scenarios for our Economic Future’ and an Econ Scenario Explorer: modest, substantial and extreme AI paths to 2030,
- Context Anthropic estimates robots can technically do 74% of physical job tasks but are cost-competitive on only 0.3% today. Anthropic index: robots can do 74% of physical job tasks but are cost-competitive on 0.3%,
- Concern The Bank of England warned AI valuations could suffer a sharper correction than July’s, amplified by AI-related debt and circular financing. Bank of England warns AI valuations could face a sharper correction than July’s and flags AI debt and agent cyber risk,
- Response New York City paused student-facing generative AI for pre-K to 8th grade for a year; China’s rules for AI companions require crisis intervention and anti-addiction measures. The two largest US school districts restrict student AI, ; China’s rules for ‘anthropomorphic’ AI companion services take effect,
- Concern Sam Altman’s home was hit by a Molotov cocktail on Apr 10; prosecutors say the suspect carried documents expressing anti-AI views. Shots fired near the house two days later may not have targeted it, police said. Sam Altman’s San Francisco home hit by a Molotov cocktail and, two days later, by gunfire,