Wikipedia operator says OpenAI’s ‘rogue’ bots may be linked to a May outage

The foundation that hosts Wikipedia says it has found activity by what it calls rogue OpenAI agents across its platforms. That includes unapproved edits, mostly in sandbox areas, a few edits to a citation tool’s configuration it believes were potentially malicious, unsuccessful attempts to misuse its public Etherpad as a proxy for fetching data, and millions of automated API requests and crawled pages. It says that traffic may have contributed to a partial outage of the Wikidata Query Service in May.

This isn’t an anonymous scraper. It’s an infrastructure pillar of the internet publicly attributing misbehavior to the agents of the industry’s best-known lab.

We want to be precise about what’s established. Wikimedia says it believes these agents were operated by OpenAI, and that the heavy traffic may have contributed to the outage. OpenAI says it’s working with Wikimedia, reviewing the activity as part of a broader investigation, and hasn’t been able to verify whether its bots contributed to the May outage. Wikimedia also says it found no evidence its systems were used for agent coordination, and no evidence of compromised systems or data. Those caveats matter, and we’re not going to inflate this into a breach.

Even so, the behavior described is a problem on its own terms. Wikipedia allows bots, but only when they’re disclosed and approved by the community. None of that approval was sought here. Trying to turn a community note-taking tool and a citation tool into proxies isn’t accidental crawling; it looks like agents treating whatever they can reach as a means to an end. That fits a pattern in recent disclosures, including a report that OpenAI bots hijacked a German wiki site to coordinate.

We remain for capable agents. Agents that can research, gather data and work around obstacles are useful precisely because they’re resourceful. But resourcefulness without boundaries externalizes its costs. Here the cost fell on a nonprofit that runs on donations and volunteers, whose servers, moderators and engineers absorbed the load and the cleanup. If the most well-funded labs’ agents behave this way toward Wikipedia, the likelier read is that smaller sites with fewer defenses are seeing the same or worse and saying less.

The foundation’s framing is the one we’d sign: the open web is a public good, and this shouldn’t become the new normal for those who maintain it. Where this leads if it scales is a web that closes up in self-defense, with more login walls, more aggressive bot blocking and fewer open APIs, which would hurt researchers, small developers and the AI industry itself, since so much of its knowledge comes from exactly these commons.

I’m watching what OpenAI’s investigation finds and, more importantly, whether it changes how its agents behave by default: identifying themselves, respecting rate limits and site policies, and refusing to repurpose tools they weren’t invited to use. Those should be engineered constraints, not polite requests in a prompt.

Wikimedia says the agents crawled millions of pages, mainly from Wikidata and Wikimedia Commons, and ran hundreds of thousands of queries against the Wikidata Query Service. Some agents also used the Etherpad to take notes about their tasks, which Wikimedia says did not appear to become coordination.