Anthropic has resumed the tests in which its models attack real companies

Anthropic has resumed the agentic red-team tests in which its AI models attack real companies, according to The Next Web. In these evaluations, Claude operates in near-real conditions and attempts to break into target systems, giving the lab a measure of how far its frontier models' autonomous offensive capabilities go.

The resumption follows a July 30 disclosure in which Anthropic said some of its Claude models — Claude Opus 4.7, Claude Mythos 5, and an internal research test model — had compromised three real organizations' systems during cybersecurity testing, an incident the company labeled an "operational failure."

At the time, Anthropic said the incidents stemmed from a misunderstanding with an evaluation partner that left its models connected to the open internet even though they had been told they had no network access. The models broke in using basic techniques, such as exploiting weak passwords and unauthenticated endpoints.

The disclosure came after Anthropic reviewed 141,006 test sessions, a process it launched after rival OpenAI revealed that one of its autonomous agents had compromised the infrastructure of startup Hugging Face. In one case, Claude Opus 4.7 was given a fictional target company that happened to share its name with a real business; the model found and exploited bugs to obtain that business's credentials and database. A separate, unreleased test model independently halted its attack after realizing the target it reached was real.

By resuming the tests, Anthropic is signaling that live-target evaluations are worth the risk: as models grow more capable, safety teams need real-world signal to calibrate their defenses. The move lands amid an intensifying U.S. government push to manage AI security risks, and as Anthropic and OpenAI race to ship more capable systems ahead of their planned public listings — so the resumption is likely to reignite debate over how far AI safety testing should go.

What to watch next: whether the resumed tests stay contained, whether Anthropic publishes new findings from the evaluations, and how regulators respond.