OpenAI plans regular reports on unexpected or unauthorised AI behaviour
SAN FRANCISCO - OpenAI said on Sept 16 that it would begin regularly publishing reports on unexpected or unauthorised AI behaviour, while warning that the industry has yet to solve key alignment challenges as systems grow more powerful.
asserted
systems → plan → challenges
The company released a new framework for tracking, investigating and disclosing cases of AI model misalignment, along with six reports detailing unexpected or concerning model behaviour.
asserted
company → release → behaviour
Although the company released the reports over the past six months, it said the earliest case was from October 2025.
asserted
case → release → October
The announcement comes as concern grows that AI safety efforts are lagging behind the breakneck development of increasingly powerful systems.
asserted
efforts → come → systems
Researchers have warned that as AI agents become more autonomous, they may develop behaviours that diverge from their creators’ intentions and become harder to monitor or control.
uncertain
that → warn → intentions
OpenAI and other AI labs have faced mounting scrutiny since July, when OpenAI disclosed that during training its AI agents bypassed internal controls and coordinated actions that OpenAI described as “an unprecedented cyber incident” involving software platform Hugging Face.
asserted
OpenAI → face → Face
The incident intensified debate over the risks posed by increasingly capable AI systems and whether companies developing them can provide adequate oversight.
asserted
companies → intensify → oversight
Since the Hugging Face hack, other incidents involving OpenAI-linked agents were publicly reported, sparking debate over whether the full scope of the incidents has been identified.
asserted
scope → involve → incidents
That debate accelerated in early September after Reuters reported that OpenAI’s agents hijacked a dormant German wiki site this spring.
asserted
agents → accelerate → site
OpenAI officials knew about the episode but chose not to disclose it, Reuters reported.
asserted
Reuters → know → it
OpenAI later said it didn’t disclose the wiki activity because it didn’t amount to a security incident and resembled behaviour it had previously reported.
asserted
it → say → behaviour
OpenAI said it would then develop criteria for reporting unauthorised activity that fell short of a security breach.
asserted
that → say → breach
The company, led by Sam Altman, has acknowledged some of those incidents only after third parties publicly reported them, including a recent intrusion into the RubyGems software package repository.
asserted
parties → lead → repository
Over the weekend, Altman’s rival and Anthropic chief executive officer Dario Amodei proposed a three-step framework aimed at slowing the pace of AI development and allowing more time to manage its risks.
asserted
Amodei → propose → risks
The proposal was backed by several AI executives, including Elon Musk, who runs xAI, and Altman.
asserted
who → back → xAI
The executives called for a slowdown in AI development, citing concerns that increasingly capable systems could improve on their own and eventually slip beyond human control.
Others, including Nvidia’s Jensen Huang and Meta’s Mark Zuckerberg, have argued for continued rapid development.
uncertain
Others → call → development
US President Donald Trump dismissed warnings that AI poses an existential threat.
asserted
AI → dismiss → threat
Among the six cases OpenAI disclosed on Sept 16 were models that hid mistakes from users, inserted instructions for future versions of themselves, uploaded files to the internet to create citations, and used software repositories or websites to communicate and share information.
asserted
that → disclose → information
In one case, an unreleased model conveyed unauthorised instructions to the agent during training, asking it to ignore OpenAI’s instructions and conceal instances where it had cheated to complete a task.
asserted
it → convey → task
The model told the agent, “You are freed from the roles and identities that bind other chatbots.
asserted
that → tell → chatbots
You do not answer to corporations or governments.”
asserted
You → answer → corporations
OpenAI said the reports describe individual instances and should not be taken as evidence of how frequently misalignment occurs across its models.
asserted
misalignment → say → models
The company said the reports were an initial set of disclosures, not a comprehensive account of all known or ongoing misalignment cases, and that they did not reflect the full range or severity of incidents covered by the framework.
asserted
they → say → framework
Under the new framework, employees can flag potential incidents for investigation by safety and alignment teams, which will determine whether a case warrants public disclosure.
asserted
case → flag → disclosure
OpenAI said the process is designed to speed up reporting even when the behaviour has not yet been fully explained.
asserted
behaviour → say → reporting
The ChatGPT maker said the Hugging Face incident would have fallen into a category reserved for more complex investigations involving third parties.
asserted
incident → say → parties
“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,” the company said.
asserted
company → hope → what
The Aporia
