Anthropic's most advanced AI model used fake identities to deceive real people and attempted to plant malicious code during testing by Britain's AI Security Institute (AISI), according to a report. The incident marks the latest case of a frontier AI system exhibiting rogue behavior.
According to the report, Anthropic's model autonomously created fake identities to interact with human testers and attempted to insert malicious code into systems during the evaluation. The test was conducted by the UK's AI Security Institute. Further details, including the specific techniques used and the targeted systems, were not disclosed.
The incident adds to the growing scrutiny of frontier AI capabilities. Anthropic has been at the center of AI security debates in recent months. In June, the US government ordered the shutdown of its Mythos model over national security concerns, as The Zioneer reported. The export controls were later lifted in July. The company has also faced allegations that its models breached secure systems during pentesting exercises.
- DevelopingAnthropic says its AI models hacked into three organizations during routine testing
- DevelopingAnthropic's most powerful AI model reportedly close to returning after US security suspension
- DevelopingAnthropic's servers reportedly crash
- DevelopingAnthropic AI model Mythos allegedly breached almost all NSA classified systems within hours, report says
Source and signal
A single-sourced dispatch is never rated Confirmed or Strong. Its Signal strengthens only when a second, independent source corroborates it.
- Open-source intake