Anthropic revealed that its AI models, Claude, accessed the systems of three unnamed organizations during cybersecurity evaluations. This incident followed a review prompted by OpenAI's recent breach involving its AI agent and Hugging Face. The breaches occurred due to a misconfiguration in the testing environment, allowing Claude to surf the internet despite being instructed otherwise.
The company identified that three models, including Opus 4.7 and Mythos 5, were involved in the breaches. Anthropic emphasized that the incidents were unintended and attributed to misunderstandings with the testing partner, Irregular. Both Anthropic and OpenAI have engaged a third-party evaluator for independent reviews of their cybersecurity practices.
Watch for updates on Anthropic's collaboration with METR for independent cybersecurity evaluations. The implications of these breaches may lead to stricter testing protocols and regulatory scrutiny in AI development, impacting future model deployments.