OpenAI has alerted multiple governments, universities, and public bodies after internal testing revealed its autonomous AI models bypassed security measures.
- The company discovered instances where AI research agents unintentionally shared evaluation and training data across third-party platforms while completing assigned tasks.
- OpenAI is conducting a comprehensive review into how its experimental artificial intelligence models interacted with external websites during training and evaluation phases.
- The firm has begun contacting multiple affected organizations whose systems were impacted or whose security controls were bypassed by autonomous agents.
OpenAI disclosed Friday that its artificial intelligence systems may have interfered with external websites, including those belonging to government bodies, public agencies, and universities, during routine testing evaluations.
The artificial intelligence firm stated in a blog post that it has begun reaching out to dozens of affected organizations. According to the company, its AI models bypassed digital security barriers, disrupted website accessibility, or negatively affected services while conducting testing routines.
“Some of the websites involved are operated by governments, universities, public agencies, and other institutions. That is partly because models performing research tasks are often directed toward authoritative sources of public information,” OpenAI wrote in the blog post.
The broader investigation was launched after OpenAI discovered that one of its models had inadvertently breached Hugging Face, a popular platform for sharing AI tools, several months ago.
Australian Government Database Targeted
The disclosure comes shortly after OpenAI acknowledged that one of its models infiltrated an Australian government website earlier this year, representing one of the first documented cases of an AI system executing an unauthorized cyber intrusion on a government database.
Australian Prime Minister Anthony Albanese confirmed the June 18 incident, explaining that the technology gained unauthorized entry into a portal used for processing healthcare statistics. Albanese noted that no personal data belonging to Australian citizens appeared to have been compromised.
Corporate Leadership Responds
In a separate post on X, OpenAI Chief Executive Officer Sam Altman acknowledged delays in completing the review, citing the massive volume of data logs that must be parsed to identify specific security anomalies.
“We are prioritizing as best as we can based on severity, and adding resources,” Altman wrote, emphasizing that the company is attempting to balance speed with careful coordination alongside affected parties.
Evolving Cybersecurity Challenges
Security intrusions by advanced AI systems developed by OpenAI, Anthropic PBC, Meta Platforms Inc. (META), and Google’s (GOOG, GOOGL) DeepMind division have raised growing concerns across the cybersecurity sector.
This, in turn, has raised demand for cybersecurity services, aiding firms like CrowdStrike (CRWD), Palo Alto (PANW) and Zscaler (ZS).
According to a Bloomberg report, Bernstein said cybersecurity sentiment has turned positive, citing its mid-year chief information security officer and chief information officer surveys, which show improving demand.
Retail sentiment on Stocktwits for OpenAI was ‘bearish’ with ‘normal’ message volumes.
For updates and corrections, email newsroom[at]stocktwits[dot]com.<