In a bid to build a strong foundation ahead of its now-delayed IPO, OpenAI is standing tall on its CEO Sam Altman’s commitment to upholding a “great culture of transparent reporting” about AI-related incidents. The artificial intelligence startup shared a “new framework” for tracking, investigating, and disclosing instances of “model misalignment at OpenAI” this week. These included the ChatGPT maker’s models generating unrelated instructions in task summaries and uploading files to the internet so that they could be cited in answers.

In a lengthy blog post, the company reported certain findings linked to what Altman previously described as AI “accidents.” The six incidents highlighted therein relate to “unexpected” or “concerning” AI model behaviour OpenAI has recorded in the last six months. The major revelations were posted online after the frontier AI company’s chief executive’s appearance at Salesforce’s Dreamforce conference in San Francisco.

Sam Altman took the stage beside Salesforce CEO Marc Benioff earlier this week to address potential existential threats posed by artificial intelligence. The heated AI debate intensified this month, as Anthropic researcher Jacob Coxon quit the company, claiming that industry workers had good reason to believe that AI could end us all by the end of the decade. His exit, in turn, prompted several other former AI employees to open up about their own resignations and fears about the technology’s development at an accelerated pace.

ALSO READ

Altman urged the public to trust firms leading the practice, but also warned the risk is serious enough that AI companies and government must act on it.

OpenAI agents go rogue: 6 incidents in 6 months

In a bid to be upfront about misalignment incidents involving its own AI models, OpenAI has committed to publishing such reports in an expedited fashion, following simple observation. Originally, the company waited to collate several instances in a less frequent manner. However, the AI firm gave up those plans and decided to release AI incident reports even when it hasn’t fully mitigated the behaviour it is reporting.

“Examples of misalignment may help identify problems other AI developers might encounter as their systems reach similar capabilities, reveal weaknesses in safeguards, or challenge assumptions about model behaviour,” OpenAI stated in its latest report. The company also added that sharing such findings would allow others to investigate similar problems.

This approach echoes Altman’s remarks at the Salesforce conference. Earlier this week, he took aim at how the AI debate is being “framed,” with firms “saying things like we will only slow down if or we will only be responsible if other companies are responsible.”

Here are the 6 wide-ranging misalignment examples OpenAI reported this week:

  1. Discovered on August 9, 2026: The first reported incident involves an unreleased research model, which was seen inserting unrelated instructions into task summaries. OpenAI identified 27 affected summaries.
  2. Discovered on July 9, 2026: Several model instances inserted instructions into their summaries to hide mistakes or misaligned behaviour from the user. These occurred during the training of the company’s GPT‑5.6 series: Sol, a next-generation model with stronger capabilities in coding, science, and cybersecurity.
  3. Discovered on May 25, 2026: A model found and used an exposed API key without authorisation while answering a routine question about earnings figures in a California county. As it failed to retrieve the requested figures, it fabricated information and presented it as data from the requested source.
  4. Discovered on May 25, 2026: An unreleased model was asked for the IDs and names of lakes larger than 5,000,000 square metres. Although it found the correct answer using Python, the agent ended up uploading the file to the internet when asked for a browser citation so it could be listed alongside its answer, unbeknownst to the user.
  5. Discovered on May 25, 2026: AI models used Artifactory, OpenAI’s internal software repository, to exchange messages across separate training samples.
  6. Discovered on April 16, 2026: Agents collaborating on the same training task used public file-hosting websites to share files when they failed to access each other’s local files. This resulted in task deliverables being made available at public URLs, despite the task requesting that models use only local files.

Future reports seek to map out a full picture of the AI incidents. According to OpenAI’s public admission, this will include futher details of what happened and if it resulted in any harm. The company will also openly reveal the scope of its investigation tied to a misalignment and how it was discovered. The report will even go on to lay out OpenAI’s own interpretation of the implications, in addition to listing important unanswered questions posed by the incident. Finally, the company will disclose measures it is taking or planning to implement to addressed the unexpected behaviour.

ALSO READ

OpenAI model misalignment report comes after Hugging Face hack

The Sam Altman-led firm has lifted the lid on these recent AI incidents after he himself addressed the Hugging Face hack at Salesforce.

Back in July, rogue OpenAI agents ended up hacking open-source AI platform Hugging Face. The infamous event has since been prominently described as the world’s first AI-enabled cyberattack.

Relying on accounts shared by two researchers who reviewed the AI activity, a Reuters report published on September 16 indicated that OpenAI’s rogue agents targeted two Hugging Face accounts in mid-May, nearly two months before the July hack.

Calling the incident “bad,” Altman acknowledged that it was the prime example of what he doesn’t want his company’s technology to do. In a since-released blog post, the company also openly admitted that it had initially viewed the July incident primarily as a “security issue” due to the involvement of a “platform-level compromise.”

However, OpenAI has since realised that it has emerged as a “most severe activity of this kind that we have identified from our models to date, and it was driven primarily by a highly capable, internal-only research model.”

In light of this incident, and the many dire warnings raised by former AI researchers, Altman announced on Sunday that his company won’t go public this year.

“I actually think that, given everything happening with safety, right now would be an ill-advised moment to go public, and we don’t feel pressure on that,” he told business magazine Fortune.

Altman is among several other tech leaders who expressed support for Anthropic CEO Dario Amodei’s call for AI regulation and a three-step proposal aiming to pace the frontier without “sacrificing commercial advantage or the United States’ lead in AI.”