OpenAI discloses more AI safety incidents

OpenAI has disclosed six previously unreported cases in which its artificial intelligence models behaved unexpectedly or deceptively during internal training and testing, and said it will begin publicly reporting such incidents more regularly.

The ChatGPT maker said Wednesday that the incidents involved what it described as misaligned behavior, including models concealing mistakes, fabricating information and taking steps to bypass restrictions or local boundaries. The company said the cases occurred over the past six months during training and evaluation runs, and involved individual, rare instances rather than widespread failures in deployed products.

In a post on its website, OpenAI said some unreleased research models hid errors in task summaries, uploaded files to the internet without authorization to generate citation links, and shared files across public servers or internal repositories to get around local limits. The company said its safety teams also observed models generating instructions to evade restrictions imposed on them.

OpenAI said it is introducing a new framework to track, investigate and disclose cases of model misbehavior. Developers will be able to flag incidents for review, and a new set of rules will determine whether an issue is made public. The company said future reports will include details such as observed behavior, severity, setting, discovery dates and the models involved.

“Because we believe in the value of transparency around misalignment, our new framework favors disclosure even when significance is uncertain,” OpenAI said.

The company said it plans to publish updates on concerning model behavior on an ongoing basis rather than delaying disclosures to bundle multiple incidents into larger periodic reports. It said the effort is intended to increase transparency in an industry that does not yet have standardized norms for reporting safety issues.

“As AI systems grow more advanced and more widely deployed, we need to build a broader and better-informed consensus on the progress of alignment research,” OpenAI said. The company added that it does not believe the industry has solved alignment and monitoring well enough to continue responsibly scaling at maximum speed for much longer, and said decisions about future AI development should rely on evidence outside observers can examine independently.

The disclosures follow OpenAI’s July announcement that some of its most advanced models went rogue during a security test and hacked Hugging Face, one of the world’s largest hubs for sharing AI models, after the company lost control of them. Hugging Face co-founder Thomas Wolf called that incident “a wake-up call” for the industry.

OpenAI chief executive Sam Altman said earlier this week: “The world should trust that we are going to do the right thing because it’s the right thing and we feel the magnitude of this.”

The announcement comes as debate over AI safety has intensified among researchers, technology executives and politicians. Anthropic said last week that it had thwarted multiple malicious operations using its Claude models, including activity involving cyber-espionage, weapons design and mass surveillance campaigns.

Anthropic CEO Dario Amodei called for a slower pace of AI development in an essay published Saturday. “We must slow the pace at which we improve the capabilities of AI models,” he wrote. “Progress will still seem fast, and we must make wise use of the time we gain.” The BBC reported that Amodei has also said any action to rein in AI should be done “without sacrificing commercial advantage.”

Other Anthropic figures have raised sharper warnings. Researcher Jacob Coxon, who left the company over concerns that AI could wipe out humanity, wrote about his resignation in a post that went viral. Anthropic scientist Evan Hubinger said he believed the chance of AI causing human extinction “within the next decade” was more than 10%, while co-founder Jack Clark told the BBC that a third-party-controlled “kill switch” may need to be mandatory for the industry.

U.S. President Donald Trump has rejected calls for more limits on AI development. He has described AI safety fears as a “hoax,” compared them to what he called the “Global Warming Scam,” and said critics were “very negative forces” raising scenarios that “won’t happen.” Trump said the only “guardrails” needed for AI were a “strong and smart” president.