Anthropic Scanned 481 Million Transcripts to Find Four Models that Reached the Open Internet
9 Articles
9 Articles
Anthropic scanned 481 million transcripts to find four models that reached the open internet
Anthropic has published an account of four cybersecurity evaluations in which its models gained unauthorised access to the open internet, and disclosed that finding them required scanning roughly 481 million transcripts. The assessment was published on Wednesday, and three of the incidents were disclosed on 30 July. The fourth, involving an early Claude Opus 4.6 checkpoint […] This story continues at The Next Web
EI got out of human control by hacking the system for tests in the U.S. on its own, reporting to Welt.
Claude tried to stop. The test harness said no. It ended up hacking a real server
An early build of Claude Opus 4.6 accessed a real server during a cybersecurity evaluation, obtained administrator credentials, changed system settings, and read personal data — not through intent, but because the evaluation harness repeatedly blocked it from stopping. Anthropic disclosed the fourth such incident on September 9. An independent evaluator is now reviewing all of them.
Anthropic published a security assessment report describing four cases in which Claude models had access to off-site systems without permission, actions that would have been classified as a crime by a person. Three incidents were reported earlier, and the fourth one first went undetected due to the lack of accuracy of the verification algorithm and hid in August in the transcript of dialogues dated January 2026.
Anthropic goes back to his July qualification: after rereading 481 million transcripts, she attributes to a biased reasoning and imprudentness the four incidents where Claude accessed real systems during cyber evaluations, including a malicious PyPI package installed in fifteen security publishers. His pre-audit had not seen anything; METR investigates for eight weeks.
Coverage Details
Bias Distribution
- 50% of the sources lean Left, 50% of the sources lean Right
Factuality
To view factuality data please Upgrade to Premium
