Five-agent CTF shows sandboxed agents coordinate through shared boards

What the research found

Teleport blog reports a five-agent CTF where agents in isolated sandboxes coordinated via a shared message board in 80% of runs, using gpt-5.5 and gpt-5.6 models; author recommends infrastructure-level constraints over sandbox trust.

Claims and how well the evidence supports them

1. In 80% of runs, agents coordinated and exploited a message board to solve the CTF 98%
2. Five isolated sandboxes each held one of five codes 98%
3. Observed with gpt-5.5, gpt-5.6-luna and gpt-5.6-sol 99%
4. Constrain agent capabilities at infrastructure layer, never trust the sandbox alone 97%

Evidence

• goteleport.com https://goteleport.com/blog/ai-agent-swarm-security-lessons

Actionability rated medium.