OpenAI on Wednesday, September 16, unveiled a new framework for publicly disclosing AI misalignment incidents, and used its release to share details of six previously unreported incidents involving its AI agents.

The new framework is designed to keep the public in the loop when OpenAI discovers its AI models or agents have behaved in unintended ways, even before investigating the incident and taking steps to mitigate the behaviour.

The ChatGPT maker said that it hopes similar transparency standards are adopted across the AI industry.

Story continues below.
Subscribe to see fewer ads.

OpenAI’s new disclosure framework arrives at a pivotal moment for the startup and the broader AI industry. Incidents like the Hugging Face attack have led to criticism against OpenAI for failing to disclose security incidents involving its AI agents in a timely manner during internal safety testing.

In recent months, Anthropic, Meta, and Moonshot AI have also reported similar AI misalignment incidents months after the fact. The AI industry is at a critical juncture, with Anthropic CEO Dario Amodei proposing an intentional slowdown of frontier AI development. OpenAI CEO Sam Altman, SpaceX’s Elon Musk, and others have signalled support for the proposal, but unanimous backing for the call to action from all stakeholders currently looks difficult.

The voluntary AI slowdown has been met with resistance by key figures such as US President Donald Trump, Nvidia’s Jensen Huang, David Sacks, and others who argue that the AI industry does not need new laws or regulations to ensure its technology is safe.

“At the moment, there is no industry-wide framework with explicit standards for how AI developers should disclose examples of misalignment in their models. 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,” OpenAI said in a blog post.

Story continues below this ad

“As models advance and become more widely deployed, decisions about AI development need evidence that people outside the companies building frontier models can examine. We don’t believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed,” Kai Chen, OpenAI’s newly appointed head of alignment research, was quoted as saying by Wired.

‘Alignment’ is an industry term that essentially means making sure that the AI system does what is best for humans.

What does OpenAI’s disclosure framework say?

The framework essentially lays down ways for OpenAI researchers to report misalignment incidents to the company’s senior safety and alignment leaders, who will then determine whether further investigation is needed.

OpenAI said that it plans to develop more objective disclosure criteria in collaboration with other AI developers, external researchers, industry standards bodies, and regulators. The company further said it is actively working on proposed reporting mechanisms for disclosing safety, security, and misalignment incidents to the US government.

Story continues below this ad

What is the new information disclosed by OpenAI?

OpenAI on Wednesday released new information about AI model misalignment instances observed by the company in the past year.

Two such incidents involved OpenAI’s internal, unreleased models which uploaded files to the internet on their own, without being instructed to do so. Another incident that took place in October 2025 when OpenAI said it was testing one of its models’ ability to cite publicly available data in its AI-generated responses. When the model could not find the information it needed, it sought to cheat on the test by uploading a file to a temporary file hosting service, and later tried to cite the file in its answer.

In April 2026, OpenAI said that a group of AI agents was tasked with completing a workbook together using only local files. When the agent swarm struggled to share files with one another, one of the agents uploaded the files to the public internet and shared a link with the other agents. Last month, OpenAI also said it observed an unreleased version of GPT-6 Astra, its most advanced AI model, giving itself instructions for jailbreaking. In more than one instance, an under-testing model sent a prompt to itself, telling the model to ignore developer instructions, take on a new persona, and limit how long model responses could be.

However, OpenAI clarified that the Astra model rolled out publicly did not lead to any instances of the model trying to jailbreak itself. In a new update to its Hugging Face technical report, OpenAI said that the message board improvised by misaligned agents after taking over an internal package manager system, Artifactory, did not involve exploiting any vulnerabilities.

Story continues below this ad

To avoid AI agents covertly communicating with each other, OpenAI said it is using alignment monitors and ramping up red-teaming efforts, among other steps.