← Topics01 essays · 02 labs
AI Security
Agent risk, browser agents, prompt injection, and the security edge cases around AI systems.
Writing
All writing →Labs & Projects
All projects →- Prompt Injection Lab
Backendless naive-vs-hardened agent simulator. Six attacker-crafted documents (direct override, exfil-via-image, fake tool boundaries, white-on-white, on-behalf-of confusion), a 10-rule detector, and a defense playbook ranked by impact.
- AI Codegen Audit
Zero-dependency CLI that flags the security footguns LLM code generators emit — eval, innerHTML, alg-confused JWTs, hardcoded secrets, weak crypto. CWE-mapped, OWASP-categorized.