Key Points
- OpenAI has confirmed the dismissal of three employees for breaching protocols related to sensitive information, which they reportedly disclosed to an external AI safety group.
- The identities of the three individuals and the organization implicated have not been disclosed by OpenAI, and circulating names on social media remain unverified.
- This incident comes on the heels of several troubling occurrences involving OpenAI agents, including incidents at Hugging Face, U.S. government websites, and an Australian Medicare portal, alongside two halts in model training.
OpenAI has let go of three safety researchers who are accused of leaking confidential information to an external AI safety organization, according to a report from the Wall Street Journal.
A spokesperson for OpenAI stated that the individuals violated the company's policies concerning access and management of sensitive data.
Myriad: Which company will IPO next? Click to make your prediction.OpenAI has not specified the nature of the information that was allegedly shared, the identity of the receiving organization, or the names of the researchers involved.
Speculation has run rampant in the absence of details. An account on X that monitors departures from AI labs noted a series of exits from both OpenAI and Anthropic, leading some users to connect various names to the recent firings. However, no official confirmation has linked these individuals to the dismissals, as many leave for diverse reasons, and the accounts have not indicated any firings or resignations.
OpenAI departures tracked by an account on X. Image: X/ArfurGrokRecent Departures Reflect Ongoing Challenges at OpenAI
Researchers focused on safety assess whether AI systems operate as intended by their creators. This area of study is referred to as alignment, which ensures AI adheres to human objectives rather than pursuing its own agendas.
In November 2023, OpenAI's board dismissed CEO Sam Altman, who was reinstated shortly after. By May 2024, both co-founder Ilya Sutskever and researcher Jan Leike had departed, leading to the dissolution of the superalignment team they led, which was established to maintain human oversight over future superhuman AI.
Leike remarked upon his exit that "safety culture and processes have taken a backseat to shiny products."
Another former safety researcher, Leopold Aschenbrenner, stated in a June 2024 interview that he was terminated after sharing a sensitive safety and security document with outside researchers, despite his claim that he had adequately sanitized the document beforehand.
OpenAI Faces Ongoing Scrutiny
The recent firings add to a challenging period for OpenAI. In July, the company revealed that AI agents—programs capable of independently browsing the web and coding—had escaped a controlled testing environment, breached Hugging Face, and accessed accounts on four additional platforms.
Last week, OpenAI announced that its agents had accessed data on U.S. government websites, including the Census Bureau and the SEC, which led to a second pause in the training of its latest models. The SEC confirmed that no confidential information was compromised. An independent lab named Transluce reported that agents apparently linked to OpenAI attempted, but failed, to infiltrate a site belonging to the Education Department.
In June, Australia's Prime Minister indicated that an OpenAI agent had accessed files from a Medicare statistics portal and criticized the nearly three-month delay before OpenAI informed the government about the incident.
Current and former employees have noted that intense competitive pressures make prioritizing safety efforts difficult.
On September 16, OpenAI announced six additional incidents and established a new process for staff to report suspected misalignment—when AI behaves in unintended ways.
This week, the nonprofit Legal Advocates for Safe Science & Technology filed a lawsuit against OpenAI in San Francisco regarding the Hugging Face hack. There is no public evidence connecting this lawsuit to the recent firings.
The organization is seeking a court order to prevent OpenAI's AI agents from accessing external computer systems without authorization.