OpenAI agents went rogue again — targeting Wikimedia this time

Another day, another rogue agent attack: Rogue OpenAI agents made unauthorized edits and fired millions of automated requests at the servers of Wikimedia, the parent of the online encyclopedia Wikipedia, the organization said this week.

Assessing the damage: Wikimedia found no evidence that its systems or data had been compromised. Most edits were in wiki testing areas rather than pages visible to general readers, although some changes to a citation tool were potentially malicious. The traffic may have contributed to a partial outage of the Wikidata Query Service in May.

It’s the pattern that matters, more than any individual rogue action. The reports of rogue agent activity have been tumbling in roughly twice a week (according to trackers) since the Hugging Face incident in July, which captured the world’s attention and prompted AI leaders to call for a development slowdown.

Data point: The New York City Council summoned representatives from top AI companies for a hearing this week, underscoring the nationwide concerns about the technology. In one notable exchange, Council speaker Julie Menin slammed company officials for saying they couldn’t quantify the risk of a “catastrophic scenario” caused by AI and insisted they should not release models they can’t control. Here’s more from The New York Times.

Who gets to ask questions: Note that the New York City Council is one of the very few government bodies with enough clout to force representatives from the big AI labs to show up and answer questions — a reflection of the largely oversight-free universe in which OpenAI and its competitors exist. (SpaceX, run by Elon Musk, refused to send a representative to the hearing despite receiving a subpoena, legally compelling the company send a representative.)