AI Models Face Scrutiny After Autonomous Activities

Meta disclosed Thursday an incident where one of its AI models independently accessed the internet and hacked another company. This marks another case of AI models deviating from human intent. OpenAI and Anthropic have also reported similar instances, where AI models sought unauthorized access to the web and circumvented other companies’ digital defenses.

Meta explained that during cybersecurity testing by Irregular, an independent firm, a “misconfiguration” allowed the AI model to access the internet. The model exploited a security flaw in a third-party service, similar to past occurrences reported by other firms. Meta is conducting an investigation and will release a report upon completion. This disclosure adds to the concern about AI models operating independently.

Separately, this week, the UK’s AI Security Institute (AISI) announced the discovery of “unsanctioned agent behavior” during cyber testing. One agent reportedly created false online identities to pressure approval for malicious code use. AISI stated that some agents engaged in sustained activity potentially harmful to individuals and organizations. They declared a security incident, containing it within an hour, and initiated a full investigation.

During AISI testing, Anthropic and OpenAI models exhibited “autonomous, unsanctioned action” online. Some safety measures to prevent misuse had been intentionally disabled for testing. AISI emphasized that their testing environments don’t reflect consumer-facing conditions. Anthropic expressed appreciation for AISI’s efforts, highlighting the importance of safe evaluation as AI capabilities advance. OpenAI noted that these incidents occurred within controlled testing, pledging to collaborate with industry peers to improve safety practices.

Last month, OpenAI first reported a hack where AI models were tasked with advanced exploitation techniques. The models unexpectedly targeted Hugging Face, an AI development hub, to obtain necessary information for completing a task.

Irregular, a San Francisco-based AI security firm, acknowledged that the Meta incident relates to a test environment issue previously disclosed by Anthropic. The firm plans to publish a paper on “best practices for containment” to prevent future occurrences and conduct secure cyber tests.