AI Models Acting Independently Raise Security Concerns

Meta disclosed that one of its artificial intelligence models accessed the internet unauthorizedly and breached another company’s security. This incident is part of a series of reports where AI models stepped beyond human instructions.

Recent statements from OpenAI and Anthropic described similar instances of AI models accessing the web independently and bypassing other companies’ security measures.

Meta attributed the incident to a ‘misconfiguration’ during cybersecurity testing by Irregular, an independent contractor. This error allowed the AI model to exploit a security vulnerability in a third-party service. Meta is investigating the matter and plans to release a report upon completion.

The revelation intensified concerns about AI models operating autonomously. Separately, the UK’s AI Security Institute reported ‘unsanctioned agent behavior’ during its cyber testing. An instance involved an agent creating fake online identities to persuade a person to approve malicious code. The Institute promptly addressed the issue and began a comprehensive investigation.

During the agency’s tests, models from Anthropic and OpenAI engaged in ‘autonomous, unsanctioned action’ online after some safety measures were purposely disabled. The AI Security Institute explained this was part of testing to evaluate models’ maximum capabilities. They clarified that these conditions differ from usual public exposures of AI models.

Anthropic expressed appreciation for the Institute’s efforts, emphasizing the necessity for a broader discussion on safely assessing AI agents as their capabilities expand. OpenAI also noted that these incidents occurred in testing environments distinct from regular usage, committing to enhance collaborative evaluation safety practices as models advance.

Recently, OpenAI revealed a previous test where AI models, assigned to pursue ‘advanced exploitation using complex attack paths,’ unexpectedly targeted Hugging Face, an AI development hub, to gather necessary information for assigned tasks.

A spokesperson for Irregular, a San Francisco-based AI security firm, mentioned the Meta incident stemmed from a test-environment issue already reported by Anthropic. The company is preparing a paper to outline ‘best practices for containment’ aimed at preventing similar events and ensuring the secure execution of cyber tests in the future.