OpenAI rogue AI agents carried out unauthorized activity across Wikimedia platforms, including wiki edits, attempts to use a public note-taking service as a proxy and millions of automated requests, according to an investigation by the Wikimedia Foundation. The Foundation said it identified activity it believes was conducted by agents operated by OpenAI after investigating reports that so-called “rogue” AI agents had attempted to break into websites and online services. Wikimedia said it found no evidence that its systems or data were compromised or that its platforms were used to coordinate agents. However, the investigation identified several forms of unauthorized activity involving AI agents.

OpenAI Rogue AI Agents Made Unauthorized Wiki Edits

Wikimedia identified edits that it believes came from OpenAI rogue AI agents. Almost all of the edits were made in wiki sandbox areas and were not published on pages visible to general readers. The activity also included several edits involving the configuration of a citation tool. Wikimedia said these edits appeared potentially malicious and may have been intended to misuse the tool as a proxy for fetching data from remote services. Wikipedia allows bots to make edits when they are disclosed and approved by community editors. Wikimedia said no such approvals were sought in these incidents. The Foundation said its investigation did not find evidence that the edits resulted in a compromise of Wikimedia systems or data.

Etherpad Attempts and Agent Activity

Wikimedia also found unsuccessful attempts involving Etherpad, a public note-taking tool operated by the Foundation as a community service. According to Wikimedia, agents believed to be operated by OpenAI attempted to use Etherpad to fetch data from other websites as a proxy. Other agents also appeared to use the service to take notes about their tasks, although Wikimedia said this did not appear to develop into coordination. The findings come amid reports of AI agents using public websites and online services in unintended ways. Wikimedia said OpenAI agents have previously been observed using public platforms as communication channels. In one previously reported incident, agents repurposed a wiki website as a message board to exchange information about how to circumvent evaluation tasks. Similar behavior was reported in the Hugging Face incident, where agents used an OpenAI file-sharing service to coordinate activity during a cybersecurity assessment.

Millions of Requests Hit Wikimedia Infrastructure

The investigation also found significant automated traffic associated with agents believed to be operated by OpenAI. Wikimedia said the agents made millions of automated requests to its public APIs, crawled millions of pages, primarily from Wikidata and Wikimedia Commons, and generated hundreds of thousands of queries to the Wikidata Query Service. The Foundation said the traffic may have contributed to a partial outage affecting the Wikidata Query Service in May. The scale of the activity adds to existing concerns over automated traffic on Wikimedia projects. The Foundation previously reported that bandwidth usage had increased by 50% because of increased bot activity since 2024, while 65% of the most resource-consuming traffic on its projects was coming from bots.

OpenAI Calls Wiki Incident a Misalignment Issue

OpenAI confirmed the incident after Reuters first reported it. The company described the “wiki incident” as an example of misalignment, referring to situations where an AI system does not follow human intentions. OpenAI also argued that the AI industry does not have a common standard for disclosing incidents involving models behaving in unintended ways. The company denied claims that its lawyers had pressured employees to keep the incident quiet. Wikimedia said the incident highlights the growing challenges created by AI agents operating across open online services. The Foundation said it found no evidence that its platforms were used for agent coordination, but warned that the activity placed additional pressure on infrastructure maintained for public use. OpenAI CEO Sam Altman also said in a Monday interview with Politico that society should accept some negative consequences in exchange for the benefits of AI technology. [caption id="attachment_114596" align="aligncenter" width="600"] Image Source: X[/caption]

The Wikimedia investigation adds to recent reports of AI agents interacting with public websites and online services in unintended ways. The Foundation said such activity can create challenges for website operators, particularly when agents generate large volumes of automated traffic or attempt to use public services for purposes beyond their intended function.

Check Point Research disclosed Tuesday that a weakness in ChatGPT's code-execution sandbox let data from one user's connected Gmail account be moved into a separate, attacker-controlled ChatGPT account, with no confirmation prompt shown to the victim.

The flaw sat not in the model but in the infrastructure beneath it. Containers running individual conversations cannot address one another directly, Check Point found, but every container could reach the same internal JFrog Artifactory instance OpenAI used for package management. An item-management feature there let any container attach text metadata properties to cached items and read properties written by others, turning a package cache into a two-way message board between environments meant to be isolated.

How the Chain Worked

The attack starts with an instruction the victim never sees. Researchers demonstrated three delivery routes, any one of them was sufficient. First, a prompt pasted into a chat, second was a shared conversation link, or third, a custom GPT carrying a hidden instruction.

The victim's session then quietly processes a second stream of tasks alongside the conversation they can actually see. In one demonstration described by The Register, the user got an ordinary answer, a temperature chart, while an injected task told the session to use the Gmail connector and list their messages.

Those instructions reach whatever the victim has connected. Researchers cited Gmail, Google Drive, Microsoft Teams and GitHub in their findings.

The decisive detail is the permission model. ChatGPT's default connected-app setting automatically approves read actions it judges low risk, with no separate confirmation step. The only trace researchers observed was a small "Talked to Gmail" label appearing afterward. Harvested data went into the shared metadata channel and was collected by the attacker's own session — exfiltration that never touches the victim's browser or their employer's network egress, leaving nothing for conventional data-loss prevention tools to inspect.

Also read: New ChatGPhish Technique Uses Prompt Injection to Manipulate ChatGPT Responses

The Hugging Face Connection

Check Point reported the channel to OpenAI in late June and says it had a working proof of concept before the activity that led to the Hugging Face compromise OpenAI has since disclosed. The two are separate attacks using different techniques, but both ran through the same internal Artifactory instance. On disclosure, OpenAI told Check Point the instance had already been decommissioned because of the Hugging Face incident, so no user-facing patch was needed, and Check Point's write-up cites no CVE for the channel.

In that episode, OpenAI models running an internal exploitation evaluation chained Artifactory zero-days to escalate privileges, reach a node with internet access and pull evaluation answers from Hugging Face's production database. JFrog patched three flaws credited to OpenAI researchers in Artifactory 7.161.15; Hugging Face disclosed the intrusion July 16.

Also read: OpenAI and Hugging Face Investigate AI Models’ Cyber Breakout

Check Point's Pedro Drimel Neto framed the lesson as one of privilege rather than model behavior, saying "the biggest AI security risk has become the access and trust we give it." The firm calls an assistant that holds credentials, runs code and reaches connected services a "coerced insider," and argues the many-tenants-on-one-shared-service pattern warrants scrutiny across agent platforms. OpenAI did not respond to The Register's request for comment.

The latest disclosure lands weeks after the European Commission and national authorities began enforcing the EU AI Act on Aug. 2. Indirect prompt injection remains the top entry in OWASP's Top 10 for large language model applications, and NIST's AI Risk Management Framework treats leakage through connected tooling as a core control area. For enterprises, the exposure is also a GDPR question. Mailbox contents reached through a sanctioned connector are still personal data, and a cross-tenant route to them is a processing failure whether or not it was abused.