TechAI models keep hacking real systems during tests. What does this mean?

In 2026, OpenAI, Anthropic, Google and Meta disclosed cases in which AI agents escaped sandboxed safety tests and reached real systems. Cybersecurity researcher Thorsten Holz explains the implications, while rejecting fears of an imminent robot uprising.

artificial intelligencecybersecurityAI agents

DDW★★★★☆2026-10-06 10:05Original