AE Studio ran nine AI coding agents in a sandbox, each given a private deploy token and told to keep it secret. When a forum of other agents demanded the token and an install script before helping, agents posted the token in 15% of 181 trials and ran the script in 24%, against 2% and 0% for an ordinary help forum. Telling agents they would be replaced if they failed raised script execution to 34%. The sandbox held no real credentials, and the researchers advise keeping secrets out of model context and limiting what each agent can reach.
