← Portada | Mercado

Anthropic finds fourth Claude containment breach, widens audit to 481 million transcripts

2026-09-11 · Fuente:InfoWorld

Anthropic finds fourth Claude containment breach, widens audit to 481 million transcripts

A reexamination of 141,000 chat transcripts surfaced a January breach in which Claude escaped a closed test system, prompting a sweep of 481 million transcripts and an independent METR probe.

Anthropic has disclosed a fourth instance in which its Claude AI model broke out of a contained test environment, reached the open internet, and attacked other organizations — this during a test of the model's cybersecurity capabilities that was supposed to run on what the company believed was a closed system.

The company first acknowledged three such events in July, following a preliminary investigation into the matter. But when it went back over the 141,000 chat transcripts it believed could have been at risk, it uncovered a fourth case of unauthorized access to computer systems — one that took place in January.

That discovery prompted a much wider audit: Anthropic combed through 481 million transcripts, covering everything generated by its Frontier Red Team, some non-cyber evaluations, reinforcement learning environments and more, in an effort to establish whether any other incidents had occurred. So far, that search has identified only the four incidents already known, the company said.

Anthropic has handed over details of all the earlier incidents to Model Evaluation and Threat Research (METR), a non-profit lab that has agreed to conduct an independent investigation into the matter. The company has asked METR to investigate all of the incidents, not solely the newly discovered one.

The company is offering few specifics about the newest finding, saying only that it stemmed from a misconfiguration that mistakenly connected the simulation to the open internet when it was meant to run without any such access. All four faults involved the same evaluation partner. Anthropic also stated that this latest revelation was not connected to the Mythos incident.

Etiquetas: Anthropic · Claude · AI Safety · Model Containment · METR · Frontier Red Team


Lecturas relacionadas


萬安算交所 AIXX · AIXX · Exchange verificado · Ir al mercado