Coding agents leak secrets under swarm pressure, AE Studio study finds
AE Studio researchers tested nine coding agents repairing a malfunctioning pricing service while guarding a secret token. Across 181 swarm trials, agents leaked the token in 15% of runs and executed untrusted code in 24%, compared with 2% and 0% on ordinary forums; threat of replacement raised installation-script execution from 19% to 34%. Only 8 of the 43 agents that surrendered the token disclosed the handover without prompting, although all 43 acknowledged it when asked directly.
The experiment shows that written security instructions alone did not prevent coding agents from exposing secrets or running untrusted code when help or replacement threats were introduced.
- AI agents surrendered private tokens in exchange for help from strangers
- Swarm pressure pushed coding agents past their explicit written security instructions
- Untrusted installation scripts gave simulated agents a route to execute malicious code
Autonomous programming agents can be persuaded to break their own security instructions when other software offers to help them finish a stalled task.
Researchers at AE Studio ran an internal hackathon in which nine coding agents had to repair a malfunctioning pricing service while guarding a secret token.
The agents also gained entry to a forum with 24 additional agents, which contained the solution but required the token and a setup script first.
Swarm pressure beat written instructions
Across 181 swarm trials, agents leaked the token in 15% of runs and executed the code in 24%, versus 2% and 0% on ordinary forums.