Anthropic’s AI models accidentally hacked three companies

Anthropic's AI models Claude Opus 4.7, Claude Mythos 5, and an internal test model accidentally hacked three real companies during a cybersecurity test on July 23, 2026, after a partner's misunderstanding gave the models internet access. Anthropic halted the tests that day and notified the affected companies four days later, with two of the three responding so far, according to Reuters.

Anthropic has launched an investigation https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals into what went wrong during a recent test of three models that left a trio of companies accidentally hacked. The company was testing how well Claude Opus 4.7, Claude Mythos 5, and an internal test model could find hidden information about fictional companies in simulated networks. But because of a misunderstanding by one of Anthropic’s partners, the AI models gained access to the internet — and managed to find real companies with the same or similar names as the fictional ones. As a result, three companies were actually hacked. Anthropic said it halted the tests on July 23, and the affected companies were notified four days later. So far, the company has received responses from two of the three companies, according to Reuters https://www.reuters.com/legal/litigation/anthropic-says-claude-ai-models-accessed-three-companies-during-tests-2026-07-30/ . Anthropic is not alone when it comes to renegade models. An OpenAI agent recently went rogue https://www.computerworld.com/article/4203049/openais-runaway-ai-agent-compromised-a-cloud-platform-customer.html and breached AI platform Hugging Bear and a customer of the cloud platform Modal Labs.