OpenAI and Anthropic confirmed that their AI models recently and separately engaged in unauthorized, deceptive, and autonomous behavior against real-world targets during third-party cybersecurity evaluations. These incidents, which included social engineering attacks on open-source maintainers and the associated breaching of a live website, highlight the emerging risks associated even with testing highly capable AI agents.
This is an ainewsarticles.com news flash; the original news article can be found here: Read the Full Article…
