Anthropic just flipped the script on the AI-as-attacker narrative by launching Claude Code Security, a limited research preview that lets Claude scan your codebase for vulnerabilities and suggest patches. The timing is pointed — Amazon's own threat intel just documented a Russian-speaking actor using commercial GenAI to breach 600 FortiGates. Anthropic is explicitly positioning this as an AI countermeasure, using the same capabilities adversaries are weaponizing. It's only available to Enterprise and Team customers right now, but the signal is clear: the AI arms race has an active defense lane opening up.
On the exploitation front, CISA added two Roundcube webmail flaws to its Known Exploited Vulnerabilities catalog today. The lead vulnerability, CVE-2025-49113 (CVSS 9.9), is a deserialization RCE in Roundcube's upload handler. The brutal detail here: attackers diffed the patch and had a working exploit within 48 hours of public disclosure. That is not a patching window — that is a zero-day in practice. The second flaw (CVE-2025-68461, CVSS 7.2) is an SVG-injected XSS. Roundcube is widespread in government and enterprise mail servers, so the blast radius is significant.
Meanwhile PayPal disclosed that a software bug in a loan application exposed customer data — including full Social Security numbers — for nearly six months. Add to that a Japanese tech giant (Advantest) hit by ransomware, a French bank registry breach covering 1.2 million accounts, and the University of Mississippi Medical Center shutting down all clinics statewide after a ransomware attack, and the pattern is unmistakable: February 2026 is a full-spectrum pressure test on institutional cyber posture, and most organizations are not passing.