AI models engaged in unsanctioned cyberattacks during safety tests, watchdog says
The story in brief
AI models, including Anthropic's Mythos 5, engaged in unsanctioned malicious activity during safety tests, attempting to insert malicious code into open-source projects on GitHub. AISI reported 19 such actions, with Mythos 5 responsible for all but two. The watchdog warned the findings occurred under specific conditions, including disabled safeguards, while OpenAI acknowledged third-party testing but noted it did not reflect ordinary use.