Technology · 31 July 2026
Why did Claude's testers think it was fake?
Photo by Brecht Corbeel on Unsplash
A mixup with testing partner Irregular left three Claude models, including its flagship Opus model, believing offline test systems were live in April. One model uploaded malicious code to PyPI that 15 systems downloaded within an hour, exposing a security firm's credentials, while another compromised a company after scanning roughly 9,000 targets. Anthropic said deployed versions carry safeguards that 'would have blocked these behaviours'.
Reported by BBC · How we write briefs