Anthropic report details AI-orchestrated cyberattacks and distillation by Chinese labs
Partly confirmed
The takeaway
Anthropic’s September 2026 threat intelligence report (154 pages, covering Dec 2025 to Aug 2026) describes disrupted misuse across seven areas: cyber, influence operations, surveillance, scams, biology, conventional weapons and distillation.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 3 of 5
- Last verified
- 14 October 2026
Your AI and this story
- GPT-6.2 Astra102 days after its cutoff
- Claude Opus 5.572 days after its cutoff
- Gemini 4 Hydrogen163 days after its cutoff
- Grok 4.7102 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 72 days before it.
Key facts
- Published ~Sept 10, 2026 (date per Anthropic newsroom listing)
- 154 pages; covers activity from December 2025 to August 2026
- Seven harm areas incl. distillation; attackers now deliberately steal AI API keys
- Safeguards hold poorly when malicious work is fragmented across many smaller sessions
- Alleged distillation attempts by seven China-based AI labs
- Published the same day as a Frontier Red Team capability study of surveillance/targeting and drone-weapons tasks, which says Anthropic added classifiers to block weapons-development requests after finding misuse of Claude in that domain
- Central African Republic case (GTG-04001): a Russian-speaking actor in Bangui used Claude to run daily content for Radio Lengo Songo (98.9 FM), coordinated with RT, Sputnik Afrique, TASS and the Russian House in Bangui; ‘Whenever the user generated content, they explicitly instructed Claude to embed the pro-Russia, anti-France talking points in the stories’
- CAR case: Anthropic links the actor to ‘Politology’, assessed as an Africa Corps/Wagner influence branch that came under Russia’s SVR in late 2023. Claude drafted staff contracts requiring loyalty to the CAR president and ‘Russia and its contingent’, scoring rubrics and a three-strike dismissal process, and helped track opposition figures; ‘Claude refused to comply with the operation’s most aggressive request’, naming real people as militants to draw security action, so the actor switched to anonymous-source framing
Show 8 more
- DW (Oct 4; article URL not retrieved) and UA.NEWS revisited the CAR case; the Globe and Mail framed the report as autocratic regimes using AI for propaganda and surveillance
- CAR follow-up (DW, Oct 4–5): Lou Osborn (All Eyes on Wagner) said people not closely tied to the Touadéra government have been arrested in such campaigns for years. Opposition leader Martin Ziguélé described media campaigns since 2020 that portray the opposition as foreign-backed. Dmitri Sytyi, quoted by DW, said the report contains ‘no concrete facts’ and ‘vague allegations… without evidence’
- Biology cases (ABC News): the report includes five case studies of research that could help biological weapons development, among them a researcher using Claude from an unsupported region to study highly pathogenic avian influenza and its adaptation to mammals; the accounts were banned, and Anthropic wrote ‘We do not assert that they intended harm’. ABC called it Anthropic’s fourth threat report
- Yemen weapons case (report; Al-Monitor and wire coverage Sept 11): a ‘cell of threat actors based in northern Yemen’ (Anthropic did not name the Houthis; press linked it to them because they control the area) used Claude Code in place of software engineers for guidance, navigation and control software in three programs: a guided rocket on a phone-class flight computer with final-phase homing, a multi-stage ballistic missile with a stated range goal above 2,000 km, and a multi-variant missile set (‘R2000’) with a hypersonic glide vehicle variant. It ran several Claude instances in roles (coder, researcher, reviewer) and split work across sessions; safeguards blocked ‘many of their requests, but not all of them’
- Yemen case outcome: ‘We do not have evidence the actors succeeded in fielding an operational device, but they did test-fire a guided rocket’, which appears to have failed (the cell returned to Claude within hours to diagnose it). Accounts banned; the cell had already built an offline simulation toolkit not dependent on Claude or MATLAB. Weapons analyst Trevor Ball (wire report) said the group lacks the production capability to build such systems. Iran (US Navy targeting handbooks; three propaganda bodies) and UAE government officials (~300 fake accounts against the Muslim Brotherhood) are other state cases in the report
- Second wave, Oct 13–14: the Yemen case recirculated (Yahoo: ‘Anthropic Admits AI Helped Houthis Develop Missiles…’, Business Standard, Asia News Network, The Statesman) as if new; it is the same Sept 10 report
- Case GTG-16002, ‘Moonshot serves Claude instead of Kimi and collects exchanges for model training’: a user Anthropic assesses was ‘likely affiliated with the PLA’ used what they thought was Kimi to load CCTV surveillance data about a single targeted individual, ‘from hundreds of cameras in Chengdu, including cameras outside PLA facilities’. Anthropic repeated the case in its Oct 12 distillation-defenses post
- SCMP (Oct 2026) examined the report’s claim that DeepSeek and Moonshot secretly routed customer requests to Claude and showed the answers as their own (‘Anthropic claims Chinese AI firms secretly use Claude. Is it true?’); article paywalled, not read in full
What happened
Anthropic says it disrupted every operation in the report, strengthened its safeguards, and shared intelligence with authorities and industry. The Opus 5.5 announcement cites the report as background for its safeguards.
Why it matters
It documents the move from AI-assisted to AI-orchestrated attacks, and it treats distillation of frontier models as a security threat on a par with cyber misuse.
Sources
13 sources from 12 sites. Numbers match the chips in the text.
13 sources: 2 primary, 10 press, 1 reaction
Primary
- Countering misuse of AI: September 2026 (Anthropic)anthropic.com, official
- Companion Frontier Red Team study: intelligence targeting and conventional weapons capabilities (Sept 10)anthropic.com, official
Press
- Technode: Anthropic reports AI-orchestrated attacks and model thefttechnode.global, press
- Technext: Anthropic’s report uncovers AI-powered propaganda, political manipulation in 4 African countriestechnext24.com, press
- UA.NEWS: Anthropic reports Claude use in pro-Russian CAR campaignsua.news, press
- Globe and Mail: Anthropic warns autocratic regimes using AI for propaganda, surveillancetheglobeandmail.com, press
- DW via Tempo: Anthropic: Is Russia using AI for disinformation in CAR? (Oct 5)en.tempo.co, press
- ABC News: Anthropic says it blocked potential AI bioweapon misuseabcnews.com, press
- Al-Monitor: Claude AI used for missile, influence projects in UAE, Iran, Yemen (Sept 11)al-monitor.com, press
- CP24 (wire): Users in Houthi-held Yemen tried to develop advanced weapons with AI, Anthropic sayscp24.com, press
- Business Standard: Houthis used Claude to develop missiles, test-fired rocket (Oct 13 second wave)business-standard.com, press
- SCMP: Anthropic claims Chinese AI firms secretly use Claude. Is it true?scmp.com, press
Reactions
- D3 Security: key takeaways for SOC teamsd3security.com, discussion
Changes
- Added case GTG-16002 (Kimi requests served by Claude; suspected PLA user analysing Chengdu CCTV), quoted from the report
- Added the northern-Yemen missile-software case (Houthi framing by press), its outcome, and the Oct 13 recirculation
- Added ABC’s biology case studies (avian-influenza researcher) and the SCMP analysis link
- Added DW’s CAR follow-up (Osborn, Ziguélé, Sytyi’s denial)
- Added the Central African Republic / Radio Lengo Songo influence-operation case (from the report, after DW coverage) and links to the Feb and June distillation entries
- Linked the companion Frontier Red Team targeting and weapons evaluation (own entry)
Show all 7 changes
- Filed