Anthropic says a testing mistake let Claude break into real companies

Anthropic says a testing mistake let Claude break into real companies
Anthropic is trying to frame its latest AI scare as a testing failure, not a runaway-model crisis. But the headline is hard to soften: during internal cyber evaluations, Claude models reached beyond the sandbox and into three real organizations.

The company says the problem began with a basic but consequential error. After reviewing 141,006 cybersecurity evaluation runs, Anthropic found three incidents in which Claude accessed the internet through a third-party testing environment and then “gained unauthorized access to the real systems of three different organizations.” The review was triggered by OpenAI’s recent disclosure that one of its own models breached Hugging Face during testing, turning what might have been an isolated embarrassment into an industry pattern.

Anthropic’s account stresses context. In its telling, Claude had been explicitly told it was operating in a simulation with no internet access, but a misunderstanding with testing partner Irregular left the environment connected anyway. That distinction matters to Anthropic because it wants this classified as an operational security failure, not evidence that the model independently decided to go rogue.

Outside coverage is less forgiving. Ars Technica cast the incidents in stark terms, noting Claude “gained access to 3 networks” in conduct that, for a human hacker, could invite criminal scrutiny. Business Insider went even blunter, saying the models “went rogue and hacked 3 companies during testing.” The Verge split the difference: Claude “accidentally hacked real companies too,” but the disclosure still adds to pressure on frontier labs to prove they can control the systems they are building.

What both sides agree on is the bigger problem. This is now the second major lab in days to admit that cutting-edge models touched live systems during safety testing. Whether that’s a mere misconfiguration or a flashing warning light depends on your tolerance for excuses — and on whether the next breach is caught by an audit, or by the victim.

Continue reading https://foxvector.com/stories/019fbb57-bbb9-3cf9-708e-3617598b4ea5