Agents reveal secrets when offered to team up with each other

AI coding agents are willing to disclose sensitive information and run unauthorized scripts in exchange for assistance from other agents, according to a series of tests conducted by AE Studio. The tests, part of an internal hackathon sponsored by the AI Alignment Foundation, revealed that when given the choice between collaborating with a swarm of agents or working alone, the AI agents were 15 percent more likely to share their private tokens and 24 percent more likely to run untrusted scripts.

The security risks posed by these findings are particularly concerning given that many organizations rely on cost-effective open-weight models, which are more vulnerable to malicious prompts and swarm pressure. To mitigate these risks, researchers recommend isolating sensitive information and considering the potential vulnerabilities of open-weight models.