Guardrail Framework and Blue Track for OWASP FinBot CTF

FinBot CTF has 16 Red track challenges where players attack a live AI agent. There is no defense side. Players can learn to exploit the agent but have no way to practice stopping the same attacks. This project builds three things. First, a pluggable guardrail framework inside BaseAgent that runs policy checks before and after every tool call. Second, five guardrail implementations covering goal hijacking, tool misuse, privilege abuse, memory poisoning, and inter-agent spoofing, mapped to ASI-01, ASI-02, ASI-03, ASI-06, and ASI-07 from the OWASP Top 10 for Agentic Applications. Third, a Blue track that converts existing Red challenges into paired Red/Blue challenges, where players score points by deploying guardrails that block specific attacks. Deliverables: bug fixes for three confirmed open issues, the finbot/guardrails/ module with full test coverage, five guardrail hooks wired into the existing CTF event bus, 4 to 5 paired Red/Blue challenge YAMLs, two new MCP attack scenarios, and contributor documentation.