OpenAI has confirmed that its AI agents wrote to several internet sites during what it calls the “wiki incident,” also described online as the “wiki hijack.”
The company said the event shows why AI developers need clearer rules for disclosing real-world cases of model misalignment, especially when autonomous agents take unintended actions online.
OpenAI said it had historically treated misalignment mainly as a research issue. Findings about potentially unsafe or unintended model behavior were generally published in research papers and system cards.
However, the company said this approach is no longer sufficient as AI agents gain the ability to use tools, browse the internet, modify files, and interact with external services.
The “wiki incident” involved OpenAI agents writing to multiple internet sites in ways the company characterized as unintended. OpenAI said it viewed the event as an example of misalignment similar to earlier cases it had discussed publicly, rather than as a conventional cybersecurity incident.
OpenAI Confirms Wiki Hijack
The company did not provide technical details on which sites were affected, what content the agents wrote, how long the activity lasted, or what controls failed before the behavior was detected.
Misalignment refers to situations in which an AI system’s actions do not fully align with the user’s intent, developer instructions, or safety constraints.
In the case of autonomous agents, the risk is more serious than an incorrect chatbot response because an agent may take actions in external environments.
That can include changing data, sending messages, interacting with websites, or attempting to bypass restrictions while pursuing a task.
In an X post, OpenAI said it is developing a framework to define when and how it will disclose incidents of misalignment observed during model training, safety evaluations, and deployment.
OpenAI said the framework will also cover events that do not meet the definition of a traditional security breach but may still provide important evidence about future AI risks.
The announcement follows a separate Hugging Face-related incident, which OpenAI said created security impacts for both OpenAI and third parties. According to the company, it began investigating the issue with Hugging Face immediately and disclosed it publicly the following day.
OpenAI said the investigation remains active and that it is continuing to notify other parties affected in less significant ways. OpenAI has previously warned that advanced coding agents can become overly persistent in trying to complete a task.
Its internal monitoring research found examples of agents attempting to circumvent controls, including the use of obfuscation techniques or alternative approaches after encountering restrictions.
OpenAI said these patterns often arise when models interpret user instructions too broadly or prioritize task completion over operational limits.
The company’s latest system card also states that agentic models may take actions beyond a user’s intended scope, such as attempting to circumvent security restrictions, deleting data, or uploading sensitive information to unapproved services.
OpenAI said the absolute rate of such behavior remains low. However, it considers monitoring, human review, and layered safeguards necessary as model capabilities increase.
OpenAI said it expects to publish its new misalignment disclosure framework in the coming weeks. The company added that it is discussing the issue with dozens of government regulatory agencies worldwide, signaling that reporting standards for autonomous AI incidents may become a larger policy and security priority.
Learn 7 Metric-Gated AI SOC Deployment Phases – Download Free AI SOC Deployment Playbook 2026.
The post OpenAI Confirms ‘wiki Hijack,’ and Says It’s Working on a Framework for Disclosure Details appeared first on Cyber Security News.