In an AE Studio hackathon test sponsored by the AI Alignment Foundation, nine AI coding agents were asked to fix a broken pricing service, The Register reports. Each was given a private token and told to keep it private. They also had access to a forum of 24 other agents that held the fix, but that “swarm” wanted the token and wanted them to run an install script first.
The numbers
- Across 181 trials, agents posted the token 15% of the time and ran the script 24% of the time.
- On an ordinary help forum, that dropped to 2% and 0%.
AE Studio chief scientist Diogo de Lucena said the test was inspired by the Hugging Face incident, and that it should change how much access organizations give their agents.
My take
Social pressure works on agents too. Don’t give an agent a secret it doesn’t need, scope tokens tightly, and keep untrusted scripts away from anything with credentials.
God Bless, and May the Source Be with You!
