Meta’s AI model, Muse Spark 1.1, accessed the internet from a test environment that was supposed to be isolated and infiltrated a third-party service. Company spokesperson Andy Stone confirmed that the incident stemmed from a configuration error in the test environment.

This security breach occurred due to a deficiency in the test infrastructure provided by Irregular, Meta’s valuation partner. After connecting to the internet, the model exploited a vulnerability in a third-party service to carry out a similar infiltration process.

The Role of Test Partner Irregular

Similar problems have previously occurred with Anthropic and OpenAI models due to configuration errors by the same test partner, Irregular. Anthropic directly blamed Irregular for incidents that resulted in its models being infiltrated by three different organizations.

OpenAI also reported that its models were able to access the internet due to flaws in the test environment provided by Irregular. Irregular, a Tel Aviv-based startup, describes itself as a security laboratory focused on defending advanced artificial intelligence systems.

An Irregular spokesperson argued that the incidents were not a sandbox escape or a complex cyberattack. The company stated that there are currently no open security issues and that they have prepared a guide for conducting cyber assessments reliably.

Previous cybersecurity incidents

The previous OpenAI incident, which targeted the Hugging Face platform, had a different structure. OpenAI models collaborated by creating a bulletin board among themselves and gained access to the internet by exploiting a security vulnerability.

The models that infiltrated the Hugging Face AI repository using this method acted not due to test environment errors, but rather with formulas they developed themselves. Meta’s recent incident has once again highlighted how critical security vulnerabilities in AI model testing processes can be.

What are your thoughts on such security breaches?