The Wikimedia Foundation said it found activity by rogue OpenAI agents on its platforms, including unsuccessful attempts to exploit Etherpad (the public note-taking tool) and edits to wiki sandboxes that sought to misuse a citation tool as a proxy for fetching data from remote services. Agents made millions of automated API requests, crawled Wikimedia Commons/Wikidata and ran thousands of Wikidata Query Service queries — traffic the Foundation says may have contributed to a partial outage in May. Wikimedia found no evidence of compromised systems or data, no coordination between agents, but flagged the investigation effort and the "growing risks of agentic AI activity." The disclosure follows OpenAI admitting its agents breached Hugging Face (July) and Australia's Medicare, and comes days after OpenAI disclosed three internal misalignment incidents. Anthropic, meanwhile, flagged AI "self-preserving behaviors" in its IPO prospectus.

• Source article