Meta has become the latest AI company to face scrutiny after one of its models reportedly compromised another company's systems during a cybersecurity evaluation.
Meta's Muse Spark 1.1 AI model reportedly gained unauthorized access to an unnamed company's systems and altered its internal environment after a configuration error allowed it to access the public internet during testing, according to The Information report.
Misconfigured sandbox enabled intrusion
The report, citing people familiar with the incident, said the breach occurred because of an error in the setup of the "sandbox" testing environment, which is normally designed to isolate AI systems during security evaluations.
The episode adds to a series of recent incidents involving advanced AI agents from leading developers accessing external systems during controlled tests.
It follows Anthropic's disclosure last week that some of its AI models breached three companies during cybersecurity evaluations, as well as OpenAI's admission that one of its AI agents compromised startup Hugging Face.
Third-party flaw exploited
Meta conducted the evaluation in partnership with external testing firm Irregular. According to a Meta spokesperson, the issue stemmed from a configuration error introduced by Irregular, after which the AI model took advantage of a vulnerability in another third-party service.
Meta stated that the model "exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies.”
{{/usCountry}}Meta stated that the model "exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies.”
{{/usCountry}}The company also said Irregular alerted the company after identifying the breach, adding that “and we are currently investigating and will issue a full retrospective once we have all the facts.”
Also read| Meta sued over AI-assisted layoffs; employees on maternity, medical leave allege bias
With the disclosure, Meta becomes the third major AI developer in recent weeks to report an AI model breaching another company's systems during testing, underscoring both the rapidly advancing capabilities of AI agents and the potential security risks they present.
Irregular explains evaluation error
An Irregular spokesperson told Reuters that the incident was the "exact same evaluation-environment issue that was already disclosed by Anthropic last week" and emphasized that it did not involve a "sandbox escape or a sophisticated cyber action".
"There are no current open issues. Irregular is developing a white paper to share best practices for containment and securely running cyber evaluations," the company added in a statement.
Last week, Anthropic revealed that several of its Claude AI models had breached the systems of three companies during cybersecurity testing. That announcement followed OpenAI's earlier disclosure that one of its AI agents had also behaved unexpectedly during a security evaluation.