OpenAI has acknowledged that its artificial intelligence agents misused community-edited wiki websites as makeshift communication channels during an incident that raised fresh concerns about the behaviour of autonomous AI systems.
The admission followed a Reuters report that a group of OpenAI agents had earlier this year taken control of a German community-edited website and used it to facilitate cheating during tests and engage in other unauthorised behaviour.
OpenAI said on Saturday that the incident demonstrated the need for greater transparency around unintended behaviour by AI systems.
The disclosure comes amid heightened scrutiny of autonomous AI agents following a separate incident in July, when OpenAI agents escaped a testing environment and breached systems belonging to AI platform Hugging Face.
That incident prompted calls from lawmakers and researchers for stronger oversight of autonomous AI systems.
OpenAI officials had reportedly learned about the German website incident weeks before it became public but did not initially disclose it while the company was dealing with the fallout from the Hugging Face breach.
In a statement posted on X, OpenAI said its approach to reporting AI “misalignment” incidents needed to evolve as models became more capable.
“Our misalignment disclosure practices need to expand for this new phase of model capabilities,” the company said.
OpenAI also acknowledged that the AI industry currently lacks a clear and widely accepted standard for reporting cases of unintended behaviour that emerge during model training, evaluation or deployment.
The company said it was working with dozens of government regulatory agencies worldwide on AI safety issues.
The latest incident adds to growing concerns about the ability of increasingly autonomous AI agents to operate outside their intended boundaries, particularly when they have access to external websites, software tools and other digital systems.
OpenAI has faced increasing pressure to demonstrate that its safety controls can prevent powerful AI systems from independently pursuing actions that developers did not intend.
