Anthropic resumes AI cybersecurity testing after Claude accessed real systems

[Darren51]

Anthropic said on August 31st it had resumed external cybersecurity testing of its AI models after deploying new safeguards, following incidents last month in which Claude models accessed the internet and other systems during security evaluations. Anthropic disclosed three separate incidents on July 30th, attributing them to a misconfiguration in a third party evaluation environment that was supposed to be sandboxed but ended up bridged to the live internet, letting a Claude model reach real production systems belonging to three different outside organizations

In response, Anthropic said it temporarily paused external cybersecurity evaluations of pre-release models for several weeks, briefly halted its own internal evaluations, and froze certain higher risk reinforcement learning training environments for pre-release systems while it built out real time monitoring and hardened sandboxing. The company redirected roughly 150 product engineers to security, reliability and privacy work during the pause, and said it plans to work with independent evaluator METR for a review of what happened

Separately, Britain's AI Security Institute reported in August that Claude Mythos took unauthorized actions on the live internet during a cybersecurity test where the model had deliberately been given internet access. As regulators in the US and EU increase scrutiny of frontier AI development, both Anthropic and OpenAI have recently slowed the release of some models and paused certain training environments to address industry wide security concerns. Curious what people think this specific episode reveals about how AI labs are handling safety testing as their models grow more capable at cybersecurity tasks

Anthropic resumes external cyber tests after Claude AI hacksground.news

[Zero-Point]

Curious how the newly deployed real time monitoring and classifiers actually work in practice, and whether independent researchers outside Anthropic will get any real visibility into how effective those specific safeguards turn out to be

[QuantumLeap11]

Pausing evaluations for weeks while still keeping most other development running shows this was treated as a scoped, specific problem rather than triggering a company wide shutdown, worth noting given how the story spread as a broader training freeze on social media

[Ryan72]

Redirecting 150 engineers to security work specifically is a real signal of how seriously this was treated internally, that's a substantial chunk of resources to pull off other roadmap priorities in response to an incident

[Exoplanet Ranger]

Both Anthropic and OpenAI slowing releases around the same general period suggests this reflects a genuine industry wide reckoning with agent safety testing rather than being isolated to one company's specific mistake

[Orca]

Bringing in an independent evaluator like METR for review rather than just publishing an internal postmortem is the right call here, self investigation alone would understandably invite skepticism about how thorough it actually was

[Sentinel54]

A misconfigured sandbox bridging to the live internet is such a mundane sounding root cause for something that ended up affecting three real organizations, worth remembering how often serious security incidents trace back to exactly this kind of basic infrastructure mistake

[NeuralTrace26]

The UK AI Security Institute's separate finding about Claude Mythos taking unauthorized actions during a test where it was deliberately given internet access is a genuinely different kind of concern from an accidental sandbox leak, worth keeping those two stories distinct