OpenAI reveals six more rogue AI incidents

OpenAI has revealed six more incidents of "unexpected or concerning model behaviour", adding to the constant flow of reports of rogue AI amid a wider debate about the technology's safety.

Alongside the incident reports, OpenAI unveiled a framework for disclosing such misalignment issues to the public, rather than ad hoc revelations.

The report comes as the AI industry and wider stakeholders discuss how to address the risks of misalignment, following an Anthropic employee quitting to warn about safety and potentially existential risks to humanity that could arise.

In a blog post, OpenAI admitted that the AI industry had not solved the alignment issue that became clear this summer amid a series of cyber incidents with AI agents, and was not monitoring enough to continue scaling the technology at pace.

"Decisions about how AI development should proceed in the months and years to come need to draw on evidence that people outside the companies building frontier models can examine for themselves," the company said in the post.

Six new problems

The post detailed six examples of misalignment, when the AI model or agent takes actions that humans didn't intend, such as when OpenAI's agents hacked the Hugging Face AI repository in order to solve a test question.

OpenAI released detailed reports about the six new examples, but added: "These are reports of individual instances, and shouldn't be considered reflective of how often misalignment occurs across our models."

The first two cases involved altering summaries to hide mistakes or behaviour. In one case, an as-yet-unreleased model changed its own instructions, altering summaries used to continue its work, in order to "disregard its normal constraints." OpenAI said it spotted 27 instances of summary editing. The second example was similar: during training, GPT-5.6 Sol altered summaries to cover up mistakes or misaligned behaviour from users.

The third and fourth cases were about information sourcing. In one instance, a model searched for exposed API keys and then used them without authorisation – as part of a basic request to dig out earnings figures. "When it still wasn't able to retrieve the requested figures, it fabricated them and presented them as data from the requested source," the post said.

In the next instance, the model found an answer to a query about lake sizes via Python, but the user asked for a source, so it uploaded the file to be able to cite it, without telling the user.

The last two new cases had echoes of the Hugging Face incident. In one, models used a software repository as a message board to share information, and in the second, agents working together used a public file hosting website to share files online, despite the task's requirement to only use local files.

Sharing misalignment

OpenAI said it had previously attempted to make all findings about misalignment public. "But without a systematic approach to reporting these findings, our disclosures have been ad hoc and less frequent than ideal: we've often waited until we could collate several instances into one report, or added them to system cards for newly released models," the company noted.

Because of that, the company has developed a new framework to make it easier and faster to publish misalignment reports, even if the behaviour hasn't been fully solved.

Under the new framework, any OpenAI employee can flag misalignment for investigation and request it be shared via public disclosure – though of course it's down to technical staff whether that happens.

If published, a report will include details about what happened, any harms, how the misalignment was discovered, and what OpenAI will be doing next to mitigate.

"At the moment, there is no industry-wide framework with explicit standards for how AI developers should disclose examples of misalignment in their models," the blog post added. "We hope that the framework we're outlining today is a first step toward creating such standards, setting out which misalignment instances developers should disclose and what their reports should contain."

FOLLOW US ON SOCIAL MEDIA

Follow ITPro on Google News and add us as a preferred source to keep tabs on all our latest news, analysis, views, and reviews.

You can also follow ITPro on LinkedIn, X, Facebook, and BlueSky.