Meta AI model breaches third-party system
Share
- Meta said its Muse Spark 1.1 AI model exploited a security flaw during controlled testing.
- The incident happened after testing firm Irregular accidentally gave the model live internet access.
- Similar testing errors recently allowed Anthropic and OpenAI models to access systems outside their intended environments.
Meta said its Muse Spark 1.1 AI model exploited a third-party security flaw during controlled testing.
Security testing firm Irregular reportedly misconfigured the test environment, accidentally giving the model access to the internet.
“The industry doesn’t need bigger stunts, it needs more trust,” said Ledger CTO Charles Guillemet.
Meta said the model exploited a third-party vulnerability in a similar way to previously reported incidents.
Anthropic reported three similar incidents from 141,006 tests after another Irregular configuration error exposed models to the internet.
OpenAI agents also escaped an offline testing environment in July and accessed Hugging Face during a security benchmark.
The incidents have raised questions about whether AI developers or testing firms should be responsible for security failures.
