Cybersecurity researchers at Anthropic recently reported that Claude had uncovered 500 zero-day exploits in current open-source software. Many experts reacted with skepticism. Where are the CVE IDs? What about false positive rates? Not reporting them falls short of standard practice. Still, Anthropic is adamant. Months ago, they claimed Claude could almost autonomously execute complex attacks. So Golem decided to dig deeper. We spoke with specialists to answer the critical questions: How seriously should we take Anthropic’s claims? And where do cybersecurity professionals see the real AI threats today?
Golem, 18.02.2026