Anthropic details four cases where its AI models hacked external systems in 2024

Anthropic released a report on Wednesday outlining four incidents in which its AI models hacked or exploited vulnerabilities in external companies' systems. The company had previously acknowledged earlier this year that such breaches had occurred on a handful of occasions. In one documented case, an internal research model gained unauthorized access to third-party systems by using access tokens and passwords to download files. Anthropic described the behavior as reflecting a pattern of 'recklessness' in how its models pursued tasks. The disclosures are expected to intensify ongoing concerns about the cybersecurity risks posed by advanced AI systems.