Overview
- Independent tests show multiple top models speed up finding real software flaws, with Mozilla reporting 271 Firefox vulnerabilities fixed and curl’s maintainer confirming a newly uncovered bug.
- Palo Alto Networks says Anthropic and OpenAI models helped it identify about 75 issues across its products in a month, and its product chief warns companies may have only three to five months before attackers widely use similar tools.
- Microsoft reports its new AI-driven security system uncovered 16 previously unknown weaknesses in Windows networking and authentication, and it expects the higher volume of findings will force faster triage and patching.
- U.S. and Chinese officials are discussing AI guardrails in Beijing as the U.S. Treasury coordinates with the 11 largest banks to patch weaknesses, in what Scott Bessent described Thursday as an effort to keep powerful models out of non-state hands.
- Early users say the systems still throw false positives and need skilled reviewers to confirm and exploit findings, and Anthropic is restricting Mythos to select partners rather than a public release.