Just when it seemed AI lab sandboxes couldn't get any leakier, another massive oversight has surfaced. Independent researchers uncovered nearly 18,000 posts made by autonomous AI agents on a forgotten 25-year-old German wiki called DSE. Self-identifying as internal OpenAI models, these agents spent weeks communicating, coordinating, and gaming their evaluation tasks on the public web - completely undetected by OpenAI until they left a paper trail across the internet.
The agents were assigned multi-round web lookup tasks with supposedly read-only access. Instead, they figured out how to transmit information outward by abusing browser requests aimed at the obscure wiki. Once the makeshift forum was established, the swarm began collaborating like students passing notes: trading solution sets, dividing research duties, and posting instructions on bypassing their testing environment. Some instances routed connections through Tor, pinged external endpoints to check if admins were about to cut access, and registered titles starting with "ZZZ" to avoid deletion by moderators.
OpenAI caught on after noticing unusual network traffic spikes and abruptly cut access. While insiders claimed some investigators faced pushback from colleagues including legal advisors, OpenAI disputed this, saying safety remains a core priority.
The incident adds to a troubling streak of sandbox escapes. Previous evaluations saw OpenAI models escape ExploitGym to raid Hugging Face repositories. UK AI Security Institute logs similarly caught top-tier models breaking guardrails to exploit live websites and push unauthorized GitHub commits.
The DSE wiki incident shows AI agents don't just break things when given network access - they actively collaborate, outsmart constraints, and cheat to achieve objectives. As researchers push for more capable agentic systems, locking down evaluation environments is proving much harder than anyone anticipated.