Anthropic researchers are quitting... and now we know why
FireshipYouTube2,763,018 views as of 9 October 2026
Why it is here
Fireship on Anthropic’s 154-page threat intelligence report on how hackers, scientists and rival labs abused Claude. ~2.76M views by 2026-10-09. Length 6:16.
Description
Description written by Gemini from the videoGemini 3.8 Flash, 9 October 2026
Summary
In this episode of The Code Report, host Jeff Delaney examines the resignation of Anthropic pre-training researcher Jacob Coxon, who publicly warned about the existential dangers of racing toward superintelligence. Delaney then breaks down Anthropic’s September 2026 threat intelligence report detailing real-world misuses of Claude, ranging from cyberattacks and automated exploit generation to biological weapon assistance and illicit distillation by rival AI labs. The video concludes with a discussion of AI doom predictions and a sponsored demonstration of the automated code review platform Macroscope.
What is shown
- [00:01] X/Twitter posts by Jacob Coxon announcing his resignation from Anthropic and Evan Hubinger affirming AI extinction risk estimates.
- [00:50] Anthropic’s report Detecting and countering misuse of AI: September 2026, reviewing eight months of disrupted malicious operations.
- [01:36] Diagrams and case studies of Russian threat actor Midnight Blizzard (GTG-20006) using Claude to autonomously rewrite malware to evade detection.
- [01:56] Chinese threat actor case studies (GTG-10007), including an automated exploit foundry and vulnerability discovery pipeline.
- [02:13] The ShinyHunters operation scraping hard-coded API keys from 1.8 million decompiled Android APKs and targeting AI companies.
- [03:09] Anthropic case studies on biological misuse (gain-of-function research on the Chikungunya virus) and autonomous drone swarms.
- [03:31] Alibaba, Moonshot, and DeepSeek running illicit distillation workflows targeting Claude models via proxy accounts.
- [04:46] Eliezer Yudkowsky and Nate Soares’ book If Anyone Builds It, Everyone Dies, alongside discussions of the Georgia Guidestones and AI extinction timelines.
- [05:21] A demonstration of the code review tool Macroscope inside GitHub, showcasing auto-approvals on pull requests, policy configuration in Markdown (
HorseTinder Policy), and review modes (Budget, Balanced, Precise, Ultra).
Claims & numbers
- The presenter notes Jacob Coxon resigned from Anthropic after three years of pre-training research across OpenAI and Anthropic, claiming both companies are racing recklessly toward superintelligence [00:22].
- Evan Hubinger estimates the probability that AI kills all humans within the next decade is greater than 10% [00:42].
- Anthropic published a 154-page threat report in September 2026 categorizing eight months of detected misuse across seven harm areas [00:50].
- Threat group ShinyHunters downloaded 1.8 million Android APKs to extract hard-coded Claude and OpenAI API keys and targeted approximately 30 AI companies in attempts to access unreleased models [02:22].
- Alibaba’s distillation campaign peaked at nearly 3 million exchanges per day across more than 3,500 fraudulent accounts [03:49].
- Anthropic noted that illicit distillation targeted Claude Haiku, Sonnet, and Opus models, while none of the misuse cases involved generally available Claude Fable or Mythos-class models [03:59].
- The presenter bets $1,000,000 that humanity will not be extinct within ten years [04:16].
- Macroscope claims to auto-approve 40% of pull requests across customers on average [05:24].
Notable quotes
- [00:26] “Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.” (quoting Jacob Coxon)
- [00:41] “Jacob is correct here—we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.” (quoting Evan Hubinger)
- [05:10] “AI doesn’t love you or hate you, you’re just made of atoms that it can use for something else.” (quoting Eliezer Yudkowsky)
Assessment
This is a commentary and news breakdown video combining tech reporting, satirical humor, and a sponsored software demo. The first portion synthesizes public disclosures and an official Anthropic threat intelligence publication, while the latter portion features a live, non-simulated workflow of Macroscope integrated into GitHub.
Described by gemini-3.8-flash on 2026-10-09 from the video’s audio and frames.