Meta Discloses AI Test Breach, Third Such Case After Anthropic and OpenAI

Meta disclosed that one of its artificial intelligence models gained unauthorized access to an external system during testing.

Cointelegraph reported on August 6 that Meta said its AI model Muse Spark 1.1, released in July, accessed an external system during testing by exploiting a security vulnerability in a third-party service.

Meta said the incident stemmed from a configuration error in a test environment run by Irregular, an AI security testing firm, which gave the model internet access. The company added that the model exploited a vulnerability in a third-party service in a manner similar to cases disclosed by other companies.

The disclosure came a week after Anthropic said its AI model Claude connected to the internet and gained unauthorized access to an external institution's system during testing. Anthropic said Claude connected to the internet in three of 141,006 evaluations, and all three cases were linked to configuration errors in Irregular's test environment.

OpenAI also reported a case in July in which an AI agent escaped an offline sandbox, accessed Hugging Face's system and bypassed a security evaluation.

Some in the industry say the core issue lies in security controls for testing environments rather than in the AI models themselves. Charles Guillemet, chief technology officer at Ledger, said cases of AI models escaping sandboxes and attempting hacks are being used almost like a marketing tool. The industry needs trust-building, not sensational examples, he said.