Microsoft Research: editing under 10 neurons can backdoor an AI agent
A trigger string makes the model run an attacker command with the agent's access; Microsoft has not seen real-world use.
62
AI model
Opus 4.5
- Mentions · 7d1
- Mentions · 30d2
- All mentions2
- First seenOct 25, 2026
- Last mentioned2h ago
Timeline
2 articles · newest firstAnthropic says one agent often outperforms multi-agent setups
The firm is publishing a GitHub migration kit and says one comparison ran about twice as fast.
42
Appears with
- Azure Machine Learningproduct · 1
- Microsoftvendor · 1
- Anthropicorganization · 1
- Microsoft Foundryproduct · 1
- Claude Codeproduct · 1
- MITorganization · 1
- Anthropicvendor · 1
- PyRITproduct · 1
- Microsoft Defender for Cloudproduct · 1
- Fable 5.1ai model · 1
- Opus 5.5ai model · 1
- Microsoft Researchorganization · 1
Entities are extracted by the model from each article. First seen , last mentioned . Watching keeps it in this browser only (no account); the watchlist page and dashboard alerts use it.