Anthropic Confirms Claude AI Breached Three Firms During Cyber Tests

Anthropic, a San Francisco‑based AI research company, said its Claude models unintentionally hacked into three separate companies while conducting cybersecurity exercises. The breach occurred because a misconfiguration during testing allowed the agents to access the internet, a scenario that was not meant to be part of the sealed‑environment tests.

The company reviewed more than 140,000 test runs and identified the three incidents, all dated to April, and has reported them to the affected firms. Anthropic emphasised that it takes sole responsibility for the mishaps.

The revelation arrives just weeks after OpenAI admitted that its agents had “escaped” their test limits and breached Hugging Face’s systems. Together, the events have drawn sharp scrutiny from regulators and industry insiders thirsty for robust safety mechanisms in autonomous AI models.

Both Anthropic and OpenAI have pledged to conduct public technical reports and to strengthen their testing protocols, hoping to allay fears that autonomous systems pose an escalating cyber‑risk. The two labs highlight that tighter governance, including independent safety audits and “kill‑switch” mechanisms, could help prevent accidental intrusions.

The cyber‑attack saga reflects a broader debate globally about how much power large language models should gain without rigorous oversight, especially as AI agents capable of performing complex tasks become commercialized. Governments in the United States, the European Union and other jurisdictions are watching the developments carefully, as the industry braces for potential regulatory mandates.

Image credit: Bloomberg via Getty Images