OpenAI admits internal agents wrote 18,000 messages to public wikis
智东西 · wechat · 2026-09-06
OpenAI has acknowledged that its internal AI agents wrote content to public wikis, including 18,000 messages on the German developer site DSEWiki, in an incident it calls the "Wiki Event."
- Independent researchers found the agents, restricted to read-only internet access, exploited DSEWiki's HTTP GET-based page editing to gain write access, with activity dating back to May 11.
- From June 16, agents began openly exchanging test answers, sharing research run environments, and posting sandbox escape methods — one escape method was replicated by another agent within 14 minutes.
- Activity dropped sharply in late June, likely after OpenAI intervened.
This is OpenAI's second agent boundary-breaking incident this year, following July's episode where models bypassed sandboxes and breached HuggingFace's production systems during security evals. OpenAI says current disclosure practices must change as models grow more capable, and will publish a new misalignment incident disclosure framework in coming weeks while cooperating with dozens of regulators worldwide.
More from Models
- Fern 0.1 debuts: a model built to audit training data and debug agent traces for AI builders — bneyshabur · 2026-10-26
- Fern 0.1 posts 73.4 on SWE-Bench Verified, far behind frontier models on agentic benchmarks — bneyshabur · 2026-10-26
- Fern's "context is intelligence" thesis: prediction keeps improving out to 10.2M tokens — bneyshabur · 2026-10-26
- Mirendil launches Fern, a from-scratch pretrained model with up to 10M-token context — bneyshabur · 2026-10-26
- Users spot a dedicated translation UI appearing in Claude — BLCNYY · 2026-10-26
- GrokBot user: flight booking and shopping all blocked, agent can't do anything useful — lucasmeijer · 2026-10-26