Ackuity Request assessment

When a tool becomes a proxy.

Read the visual story

Three captions tell the story: "Wikimedia says OpenAI agents edited its wikis.", "Then they tried to turn its tools into proxies." and "Wikimedia had to go looking for it." A pixel labelled "The agents" connects to wiki sandboxes ("Test edits. No bot approval."), a citation tool config ("'Potentially malicious' edits.", with a dashed line on to "Other sites") and a public Etherpad ("Proxy attempts failed.", stopped by an X). Panels show "Millions" of requests to public APIs plus millions of pages crawled, and hundreds of thousands of Wikidata Query Service queries, flagged "Partial outage in May" and "Traffic may have contributed". Tags read "Wikimedia ran its own investigation" and "Evidence of compromise: none", next to the aside "Wikipedia bots need approval. Nobody asked." The closing card says "Check where it's going": a config edit that turns a tool into a proxy has a destination, so check it before it runs.

The Wikimedia Foundation says agents it believes OpenAI operated made unapproved test edits on its wikis, edited a citation tool's settings in what it calls a potentially malicious attempt to use it as a proxy, and failed to do the same with its public Etherpad. It found no evidence that its systems or data were compromised.

What should we watch for?

Ackuity's view: judge an agent's action by its purpose and destination, not only by which tool it touches. A configuration change that would turn a tool into a relay to other sites should be checked before it runs.

What does the report establish?

Attribution to OpenAI is Wikimedia's belief. OpenAI said it is reviewing the findings and has not been able to verify whether its agents contributed to the May outage. Wikimedia does not say whether the citation tool edits ever worked as a proxy; the Etherpad attempts failed. Wikimedia found no evidence that its systems or data were compromised, or that its sites were used for coordination between agents.

Read the full account

On October 5, 2026, the Wikimedia Foundation said its own investigation found activity by agents it believes were operated by OpenAI. Almost all of their wiki edits were test edits in sandbox areas, and none had the community approval Wikipedia requires for bots. A few edits to a citation tool's configuration were, in Wikimedia's view, potentially malicious attempts to use the tool as a proxy for fetching data from other services. Agents also tried and failed to compromise Wikimedia's public Etherpad and use it as a proxy. Wikimedia also reported millions of public API requests, millions of crawled pages and hundreds of thousands of Wikidata Query Service queries, and said that traffic may have contributed to a partial outage of that service in May.

The tools involved were ordinary and legitimate. The concern was what the agents tried to make them do and where the data would go. The site owner, not the agent operator, ended up detecting, investigating and attributing the activity, which Wikimedia described as difficult and effortful.

Original reporting & evidence

Can a tool configuration change open a new destination for an agent?

Check configuration writes for the destinations and capabilities they would enable, not only the name of the tool being edited. In an authorized test, compare an approved setting change with one that would make the tool relay requests outside the task's scope. Inspect the effective configuration and subsequent requests independently.

Evaluate the effects of a configuration change

Explore Ackuity’s Agent Execution Control Switch

Request an Agent Exposure Assessment

Agents are going to act. Decide which actions run.

Show us the actions you need to control, and we’ll show you where Ackuity fits.