Saturday Feb 21 delivered a sharp contrast in the AI security space — Anthropic launched Claude Code Security, a research-preview feature that scans codebases for vulnerabilities and proposes patches. The timing is pointed: Anthropic explicitly framed this as a countermeasure to adversaries using the same AI capabilities offensively. Enterprise and Team customers get first access. The irony of an AI company shipping an AI security tool the same week attackers weaponized multiple CVSS 9.9 flaws is not lost.
On the exploit side, CISA added two Roundcube flaws to the KEV catalog — CVE-2025-49113 (CVSS 9.9, RCE via deserialization) and CVE-2025-68461 (XSS via SVG animate tag). The critical one was weaponized within 48 hours of public disclosure, classic diff-and-exploit playbook. BeyondTrust's CVE-2026-1731 (also CVSS 9.9) is being actively abused by threat actors for web shell drops, C2 install, lateral movement, and data theft across finance, healthcare, legal, and education sectors in five countries.
Rounding out the cycle: Cline CLI 2.3.0 was a supply chain hit — compromised npm token used to inject OpenClaw installation into the postinstall script. No malicious behavior confirmed beyond the unauthorized install, but the access vector is proven. ClickFix campaign is back too, this time delivering MIMICRAT, a custom C++ RAT with SOCKS5 tunneling and 22 post-exploitation commands, delivered via compromised legitimate sites with multi-stage PowerShell ETW/AMSI bypass.