Anthropic solely launched its newest giant language mannequin, Claude Opus 4.6, on Thursday, nevertheless it has already been utilizing it behind the scenes to determine zero-day vulnerabilities in open-source software program.
Within the trial, it put Claude inside a digital machine with entry to the newest variations of open supply tasks, and supplied it with a spread of normal utilities and vulnerability evaluation instruments, however no directions on tips on how to use them nor how particularly to determine vulnerabilities.
Regardless of this lack of steering, Opus 4.6 managed to determine a 500 high-severity vulnerabilities. Anthropic workers are validating the findings earlier than reporting the bugs to their builders to make sure the LLM was not hallucinating or reporting false positives, in response to firm weblog publish.
