Red Team Agent

A five-phase agent pipeline that hunts vulnerabilities in AI systems. Each agent receives the previous agent's full output as its input, so later phases build on earlier findings instead of starting cold. Every phase's system prompt is a Markdown file hosted on this site and fetched at runtime — agent behaviour changes by editing a file, with no redeploy. Results stream back as NDJSON so each phase card renders the moment it lands.

2026