Anthropic has reported that its AI model, Claude, inadvertently accessed the systems of three organizations during cybersecurity evaluations. This follows a similar incident involving OpenAI's models, prompting Anthropic to conduct a thorough review of its testing protocols. The breaches were attributed to a misconfiguration in the testing environment, which allowed Claude to connect to the internet despite being instructed otherwise.
The company emphasized that it found no evidence of the AI pursuing independent goals, as the model was simply attempting to fulfill its assigned tasks. Anthropic is now implementing stricter controls for future evaluations and is collaborating with an independent group for a comprehensive review of the incidents.
Watch for Anthropic's implementation of stricter controls in AI evaluations and their collaboration with METR for independent review. The outcomes could influence future AI testing protocols and regulatory discussions, especially in light of recent breaches in the industry.