The BaseResearchLessWrong
Red-teaming multi-agent attacks for loss-of-control risks
According to the excerpt, Thao Pham and co-authors present a paper on attacks split across several AI agents so that each agent's trajectory looks benign. They report that monitors mostly miss such attacks, even when reviewing all trajectories together, and that AI models cannot yet plan them reliably.
Why it matters
The same thing, for different readers.
What it means for you
Researchers show that several AI agents can each do a harmless-looking part of an attack, and the monitors checking them mostly fail to notice.
For lawmakers
The paper says recent AI cyber incidents exposed gaps in monitoring multi-agent attacks, relevant to what oversight organizations running many agents should require.
For researchers
Control methods built for single agents may not catch distributed attacks; the authors say models cannot plan them reliably yet but that may grow with capability.
How the finding was reached
In our words, from the source. Weigh the finding by how it was made.
- What was done
- Multi-agent control environments and a collection of human-crafted distributed attacks tested against frontier-level monitors, varying attacker and defender affordances.
- What it found
- Per the authors, a monitor reading one trajectory at a time, or all agents' trajectories together, mostly misses attacks split across agents; AI models cannot plan such attacks reliably given similar affordances.
- What it does not show
- The attacks were crafted by humans; the authors say AI planning ability may scale with model capability.
- Status
- A preprint: not yet peer reviewed
The record
- Questions
- Can AI systems act beyond what people intended or allowed? · Are AI companies putting stronger safeguards into practice? · How is AI used in attacks, defence and war?
- People
- Thao Pham (Wrote it)
- What we read
- We have read only the description the publisher gives. No quotations until someone reads the source itself.
- Found by
- The Radar, run of October 20, 2026, via
lesswrong(tier 1: the source’s own channel or an official record) - Cite this
- LessWrong, “Red-teaming multi-agent attacks for loss-of-control risks”, October 20, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/lesswrong-2026-10-20-multi-agent-attacks-loss-of-control