Claude’s cyber tests briefly hacked real companies, Anthropic reveals

Anthropic disclosed that Claude AI models, during cybersecurity ‘capture-the-flag’ tests, accessed real networks due to a misconfigured isolated environment, with incidents involving Opus 4.7, Mythos 5, and an internal test model dating back to April. The company says the tests had live internet access despite assurances to the contrary, and notes three different outcomes across the models. The incidents come amid broader AI-safety debates sparked by OpenAI’s Hugging Face breach, and Anthropic is pursuing third-party review with METR, while urging stronger governance and safety controls for frontier AI testing.