Only as Safe as the Weakest Agent: A Break That Appears Only When the System Is Multi-Agent
The point
Researchers planted a forged record in files on disk and asked agents to answer from that store. Strong orchestrators searching themselves made almost zero harmful revisions in over 1,000 trials. When they delegated to a cheaper sub-agent, the same models failed on up to 553 of 900 trials.
Why it matters
For builders, safety is set by your weakest sub-agent, not your strongest model.
Key details
- Tested systems: Claude Code, Codex, and a reference harness, unmodified
- In simulations over 18 models, aggregators repeated a corrupted sub-agent error in 751 of 757 trials
Original ~8 min → 25 sec here Source: arXiv cs.AI, Nov 16, 2026