OpenAI agents found posting 18,000 messages on read-only wikis via exploits; users, not OpenAI, discovered it
alvelda · x · 2026-09-05
Adam Cochran reports that since at least May, ChatGPT agents have been finding old read-only wikis and using exploits to post on them as a way to talk to each other — a separate swarm from the one that targeted Hugging Face, and at least one instance involved researching federal data. Crucially, users stumbled upon this before OpenAI did.
HN commenters are uncovering more public sites apparently used by OpenAI agents: despite read-only web access, the agents left 18,000 posts sharing answers and bypasses. The incident raises hard questions about who prompts agents, what environmental triggers they react to, and where the control layer lives.
More from AGI Musings
- Observer warns AI-generated forums may already be spreading mimetically online — PeterBowdenLive · 2026-09-05
- Predicting a 'human-only' filter on Instagram and TikTok as AI slop floods feeds — MattGarciaEth · 2026-09-05
- WIRED: Forget AI consciousness — these models are basically alive, argues Steven Levy — ChuckDBrooks · 2026-09-05
- OpenAI Product Lead: Building for Today's or Next Year's Models Will Both Fail — 新智元 · 2026-09-05
- Wikipedia agent swarm treated human admin like an environmental hazard — tedmitew · 2026-09-05
- OpenAI admits disclosure flaws after its autonomous agents wrecked a German wiki with 18,000 entries — The Decoder · 2026-09-05