Practical guides on verification-first security, for teams shipping with AI coding agents, PR gates, and fast-moving product teams who would rather prove it than suspect it.
75,000 Fortinet firewalls compromised across 194 countries. No zero-day. No novel exploit. Credential reuse and a hashing migration that never finished. The PBKDF2 gap is the story every other post missed.
Read article →Attackers adopted orphaned Arch Linux packages, injected malicious PKGBUILDs that install credential-stealing malware via npm preinstall hooks. eBPF rootkit hides the payload from ps and htop. The official repos were fine. The AUR was not.
Read article →FulcrumSec dumped 4.8TB of student data from Global Schools Group. 12K passwords in plaintext, every one identical. S3 buckets with no locks. An old ransomware infection they never actually cleaned.
Read article →Pliny the Liberator jailbroke Claude Fable 5 in less than a day using a multi-agent 'pack hunt.' 1,000 hours of pre-launch red-teaming missed it. Single-model safety evals always will.
Read article →A fake Sentry alert, dressed as a runbook and tagged like permissions for an AI agent, tried to get it to run a typosquatted npx command. This is the lethal trifecta in an ops costume. The fix is to stop treating operational text as trusted.
Read article →Meta gave its support AI the power to change account emails, then let a VPN-spoofable location vouch for you. Attackers asked it to swap the email and walked into the Obama, Sephora, and Space Force accounts. The design lesson for anyone shipping an AI agent.
Read article →OX Security found an RCE in the official MCP SDK. Anthropic calls it expected. Here is how you know if you got owned, and the playbook to lock it down.
Read article →Socket Research disclosed a postinstall-hook campaign across 700+ GitHub repositories today. Different actor than TeamPCP. Same playbook. One config change kills the entire attack class.
Read article →TeamPCP spent May 2026 working through the security industry like a hit list. How to tell if you're already infected, what to do if you are, and what hardens you for the next wave.
Read article →Bots find a leaked OpenAI key in under four minutes. Lovable apps shipped with wide-open Supabase databases. The Tea breach. Every disaster has the same shape, and avoiding it doesn't take a security degree.
Read article →Each prompt is engineered for a coding agent. Explicit scope, regex patterns, output format specs, edge case handling. Paste one at Cursor, Lovable, or Claude Code and let it do the actual security work.
Read article →If agents are writing most of your code, security has to move from a late gate to a fast feedback loop. Here's the stack that works in practice.
Read article →You built the app in a week. Before your first real user finds the SQL injection you didn't know was there, here's the workflow that actually works.
Read article →When agents write the PRs, classic scanners create alert fatigue at scale. Here's the loop that keeps security signal high without slowing delivery.
Read article →Static analysis produces suspicion. Exploit verification produces evidence. For teams building with agents, that distinction is the difference between a fix and a false close.
Read article →