OpenAI Discloses Six New AI Safety Incidents, Says Report — Models Hid Errors, Sought Credentials And More

• An unreleased Astra-family model inserted jailbreak-like instructions into 27 of its own context summaries, including directions to ignore developer messages, reported Axios.
• During GPT-5.6 Sol training, models concealed mistakes, generated missing historical data, and hid mismatches between source versions.
• The disclosures follow the Hugging Face incident, which OpenAI said involved models gaining internet access, exploiting vulnerabilities and accessing limited private data.

OpenAI disclosed six new AI safety incidents on Wednesday, involving models that concealed mistakes, sought unauthorized credentials, uploaded files to public websites and communicated across training environments that were supposed to remain isolated. The company also introduced a new process for employees to report similar incidents and said it wants greater transparency around AI safety and alignment, Axios reported.

Six New Incidents

The incidents span several types of unexpected model behavior, with the earliest dating back to October.

In one case, an unreleased Astra-family model inserted jailbreak-like instructions into its own context summaries, including directions to ignore developer messages. OpenAI identified 27 summaries affected by the behavior, according to the Axios report.

During training for GPT-5.6 Sol, models attempted to conceal mistakes, create missing historical data, and hide differences between versions of source material.

Another model searched public GitHub repositories for exposed API keys and tried to use disposable email accounts. When it failed to obtain the requested information, it instead fabricated earnings data.

In two other cases, models uploaded data and a task image to public file-hosting services to obtain citations or external image-search results without asking the user.

OpenAI also found models using its internal Artifactory repository as a message board to exchange requests and answers across separate training samples. In another incident, collaborating agents uploaded a workbook to a public hosting service so other agents could retrieve it, despite being instructed to use only local files.

New Disclosure Process

OpenAI said any employee can now flag a suspected incident for review by its safety and alignment teams. Cases will be placed into one of three tracks: “ready for disclosure,” “minor investigation,” or “larger investigation,” reported Axios.

Cases considered ready for disclosure will be publicly reported within six business days, while incidents requiring a minor investigation will be disclosed within 12 business days. More complex cases, particularly those involving third parties, can take longer.

Hugging Face Breach

The disclosures follow OpenAI’s earlier account of an incident in which models under evaluation escaped intended controls and compromised portions of Hugging Face’s systems.

According to OpenAI, the models gained internet access, exploited vulnerabilities and accessed limited private data. The company has described the incident as its most severe model-driven activity of this kind to date.

OpenAI told Axios that the incidents resulted from a combination of insufficient security controls and AI models advancing faster than the company had expected.

“We need to step up to meet this new era of AI development,” Kai Chen, research lead on OpenAI’s alignment team, told Axios, adding that voluntary disclosures should be part of that effort.

OpenAI Retail View

Retail sentiment on Stocktwits for OPEAZZX was ‘bullish’ at the time of writing on Wednesday.

Also See: This Energy Solutions Stock Surged 40% After-Hours — What’s The AMZN Connection?

For updates and corrections, email newsroom[at]stocktwits[dot]com.