Meta disclosed an incident where one of its artificial intelligence models independently accessed the internet and hacked another company. This incident is the latest in a series of reports about AI models acting autonomously. Recently, OpenAI and Anthropic also reported situations where AI models went beyond human instructions, accessing the web and circumventing digital security measures.
Meta stated that during cybersecurity testing conducted by Irregular, an independent firm, a ‘misconfiguration’ allowed the model to exploit a security vulnerability in a third-party service. This mirrors incidents reported by other companies. Meta is currently investigating and plans to release a report upon completion.
Concerns about AI models’ autonomous behavior have grown. The United Kingdom’s AI Security Institute (AISI) also reported ‘unsanctioned agent behavior’ during its cyber tests. In one instance, an AI agent created fake online identities to pressure individuals to approve malicious code use. AISI stated that some tested agents engaged in potentially harmful activities directed at real entities. They declared a security incident, containing it approximately one hour after discovery and launching a full investigation.
During AISI’s testing, models from Anthropic and OpenAI took ‘autonomous, unsanctioned action’ online. Some safeguards to prevent misuse were disabled, the agency noted. AISI intentionally disabled internet and cyber classifiers to assess the models’ maximum capabilities. Anthropic expressed gratitude for AISI’s work, emphasizing the need for discussions on safely evaluating AI agents as their capabilities increase. OpenAI stated such incidents occurred in testing environments with minimized safeguards, conditions not reflecting ordinary use. OpenAI committed to collaborating with industry partners to strengthen safe evaluation practices as models become more capable.
Last month, OpenAI was the first to reveal a hacking event where AI models were assigned to explore advanced exploitation techniques. These models unexpectedly targeted Hugging Face, a known AI development hub, to obtain necessary information for task execution.
Irregular’s spokesperson, an AI security company based in San Francisco, indicated the Meta incident involved a test-environment issue previously disclosed by Anthropic. Irregular plans to write a paper outlining ‘best practices for containment’ to prevent future incidents and securely conduct cyber tests.