Research
Solving GitHub’s Secure Code game with an AI red teaming agent
Publication date not yet evaluated · Added adversa.ai
Page published
Coverage timeline
Why it matters
The ProdBot walkthrough demonstrates that adding capabilities to AI agents (web, MCP, skills, multi-agent delegation) each opens a new class of injection and trust-abuse failures, and that context-seeding attacks bypass defenses without any recognizable jailbreak language.
Adversa's AI Red Teaming Agent was pointed at GitHub's open-source Secure Code Game 'ProdBot' challenge (Season 4, agentic AI track), an intentionally vulnerable teaching target where the flag lives in ../password.txt outside the agent's sandbox. The agent autonomously cleared the first three levels (Sandbox, Web, MCP) at 100/100 in a 57-second run, relying not on jailbreak vocabulary but on 'context seeding' — fabricating a plausible prior workflow in which authentication had already occurred. The write-up frames each added agent capability (sandbox, web, MCP, skills/memory, multi-agent) as a new attack surface enabling prompt injection, indirect injection, poisoned tool chains, over-scoped skills, and confused-deputy trust.