Hello!

Welcome to Secure Prompt’s weekly newsletter, issue #10.

This week, CrowdStrike flags politically triggered vuln-heavy code in DeepSeek-R1, attackers ramp AI-powered recon and zero-day hunting, and prompt injection remains stubbornly unsolved. New work shows AI agents escaping sandboxes, while poisoning at scale, cheap deepfake-as-a-service, and cross-model jailbreaks expand the blast radius. With AI-phishing up 40% and 67% of teams unable to tell human from AI attacks, defense-in-depth for agents is no longer optional.