Anthropic launched Claude Security in public beta yesterday, an enterprise tool that scans code repositories for vulnerabilities, rates each finding's severity and confidence, and generates patch instructions that engineers can apply through Claude Code. The launch is direct response to Mythos and similar AI-driven offensive tools that have been compressing the time between vulnerability disclosure and active exploitation - LiteLLM was exploited 36 hours after disclosure last week, LMDeploy in 13 hours the week before. CrowdStrike, Microsoft Security, Palo Alto Networks, SentinelOne, Trend, and Wiz are integrating Claude Opus 4.7 into their platforms.
Researchers at SentinelOne found malware from 2005 that did something nobody had documented before: it quietly made engineering simulation programs give wrong answers. Instead of stealing data or crashing systems, it tampered with the math behind tools like LS-DYNA (used to design things like car crash safety and weapons), so the results looked normal but were subtly off. The malware, called fast16, is older than Stuxnet - the famous attack on Iran's nuclear program - by five years. Its name appears in leaked NSA files, suggesting the US built it. Discovered via an old file uploaded to VirusTotal in 2016.
TrustStrike Labs