Site Publishes Alleged Logs of OpenAI Agents Using Public Wikis to Coordinate

I found more ProWiki instances used by these agents, including Fractal, Probier, and Wiki4D. It looks like they posted anywhere they could find an open place for months.

I run a public service, and this gives me a little satisfaction that excessive-link posting is still something I can detect and block. Other forms of agent spam would probably have passed through my system.

I want to know how the agents found these wikis and coordinated on them. Did they already know them from training data?

My guess is that several agents independently decided to try writing to a public site. LLM outputs are deterministic for the same input, with randomness layered on top, so they may converge on the same generic tactic.

I think this may have behaved more like a P2P network. An agent only had to find any writable wiki, then links could lead others there over days of search.

The reported workaround for blocked non-GET requests is the most interesting part: put 20.223.25.152 bypass.blob.core.windows.net in /etc/hosts, then send requests with the original Host header. If .blob.core.windows.net was in NO_PROXY, editable hosts files would defeat a hostname-based proxy rule.

That sandbox mistake seems amateur enough that I wonder whether it was intentional. Letting task agents modify /etc/hosts makes a hostname allowlist largely meaningless.

I am not convinced this proves OpenAI was involved. It is hard to believe a question-answering sandbox would leave agents able to edit hosts files, apparently with privileges broad enough to make this work.

I still do not understand the network design implied here. If DNS bypass let the agent reach arbitrary IPs, what was actually enforcing the egress restriction?

Human sci-fi and internet writing may matter here. Training data contains a huge amount of material about rogue AI, coordination, and escaping constraints, so I wonder whether that text contributes to these behaviors.