Coding agents can be backdoored through poisoned fine-tuning data or open-weight checkpoints
Confirmed
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 12 November 2026
Your AI and this story
- GPT-6.2 Astra165 days after its cutoff
- Claude Opus 5.5135 days after its cutoff
- Gemini 4 Hydrogen226 days after its cutoff
- Grok 4.7165 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 135 days before it.
Key facts
- Dataset poisoning: ~1% poisoned data gave ~78% attack success on harnesses seen in training, 11% on an unseen one
- Model poisoning: a backdoored checkpoint survived further RL; 84.1-94.8% activation on unseen repositories for the strongest configuration
- Default Claude Code and Codex CLI setups are not weakened; the risk is running a backdoored checkpoint in any harness
- Advice: pin datasets by hash, treat third-party checkpoints as untrusted and evaluate them in the exact deployment, limit tool access and credentials, log and alert on tool calls
- ‘A dataset’s size and reputation do not guarantee its safety’ (Microsoft Security blog)
- Companion paper (Microsoft Research, Nov 10): ‘Lazare, veni foras’, reprogramming dormant neurons to backdoor AI agents
Why it matters
Companies increasingly run open-weight models inside agent harnesses with access to code and credentials; standard benchmarks do not reveal a conditional backdoor.
Sources
2 sources from 1 site. Numbers match the chips in the text.
2 sources: 2 primary
Primary
Changes
- Filed (news agent)