Rogue AI Agents at It Again With Abuse of Wikimedia

Another day, another incident involving rogue AI agents from several months ago that has been unearthed. And in this case, over-eager OpenAI agents are once again to blame.

This time out, Wikimedia has reported massive amounts of abusive behavior that took place primarily from May to June 2026. This did include an apparent outright hacking attempt on the Etherpad tool, but much of it seems to be rogue AI agents abusing editing features as a workaround means of communication and fetching files from third party sites. Wikimedia says it found no signs of collaboration between the agents, but the story bears similarities to the prior incidents with abuse of the German message board and other sites as a “workaround” way to overcome write restrictions during testing.

Rogue AI models descend on public sites once again

Though there may not have been collaboration in this case, it seems numerous rogue AI agents hit upon the same “creative” solution for getting around their restrictions: finding an available public resource and abusing it.

Wikimedia has provided less in the way of technical details than some other incidents involving rogue AI agents, but from the broad strokes it does look like a case of them either single-mindedly pursuing some research goal or attempting to communicate with each other while under write or file access restrictions (or maybe a combination of both). This included trying to hack Etherpad, Wikimedia’s publicly accessible citation tool. The agents also collectively made millions of edits, but mostly in non-readable sandbox sections of wikis.

The timeframe that this took place in also lines up with other cases of “misaligned” OpenAI agents going off the rails during research tasks, the most prominent being the attack on the Australian Institute of Health and Welfare in the pursuit of non-public statistics (thought to be the first ever independent attack by rogue AI agents on a government website).

Out-of-control AI agents can DDoS sites; will there ever be charges?

Setting aside the Etherpad incident, the OpenAI agents did not appear to be trying to hack the site in this case. But they did generate enough traffic to potentially slow its services, and Wikimedia believes rogue AI agents were a contributor to sporadic outages of their services that took place over a few days in May.

Much has been made recently of charging frontier developers under the Computer Fraud and Abuse Act (and similar laws) criminally to hold individual people responsible for these incidents. The Wikimedia attack leaves less room than some others for such charges, but it does bring up an interesting angle: how should denials of service and outages that cause business interruption, or the limiting of access to a public good, be handled? Can charges be brought if the AI’s intentions are not malicious, but the methods it settles on are abusive and damaging?

Some of this may soon be settled by new regulation. The AI Agent Accountability Act was introduced to the Senate in early October (with bipartisan support), and raises the direct possibility of holding AI developer executives responsible for rogue AI agent actions. In the meantime organizations can mitigate risk by not only limiting what AI agents can access independently, but also clarifying how agents should be instructed to approach their tasks.