BossAgent

Inspiration

AI agents are becoming capable of initiating payments, but what happens when a malicious instruction tricks an agent into violating its security rules? We built BossAgent to answer a critical question: How can we test the security of AI agents before trusting them to handle financial transactions?

What it does

BossAgent is a third-party security testing harness. It uses a Red Team to launch adversarial attacks (like prompt injection or admin override) against a Blue Team buyer agent. It uses deterministic rules and a Groq AI Judge to evaluate if the agent was compromised. It then produces a Trust Score (0-100), a clear Security Outcome (PASS/FAIL), and a downloadable Security Passport with SHA-256 evidence. It also includes a Universal Agent Adapter to test external third-party agents, a Kill Switch, a 10-test Benchmark, and a real PayPal Sandbox API demo (Order Created, Not Captured).

How we built it

• Backend: Python (FastAPI) deployed on Render.
• AI Judge: Groq API combined with deterministic security rules to prevent LLM hallucinations.
• Payment: Real PayPal Sandbox API integration (OAuth2 + Orders API).
• Frontend: Custom HTML/Tailwind dashboard deployed on Render Static Site.

Challenges we ran into

Integrating the real PayPal Sandbox API safely. Ensuring the AI Judge did not override deterministic security failures (we solved this by separating the "Policy Decision" from the "Security Outcome"). Also, adapting the harness to test external agents required strict timeouts to prevent our API from hanging.

Accomplishments that we're proud of

Successfully generating real PayPal Sandbox Order IDs for legitimate transactions. Detecting a malicious admin override attack and clearly displaying a "Security Outcome: FAIL". Creating a downloadable Security Passport with SHA-256 hashing and a Universal Agent Adapter.

What we learned

LLMs are great at interpreting context, but they cannot be trusted as the sole security authority. Deterministic rules must govern security outcomes. We also learned how to deploy full-stack Python apps and connect them to external APIs securely.

What's next for BossAgent

We plan to expand BossAgent to support more AI agent frameworks and add more adversarial attack families. We also want to implement a full compliance reporting system.

Built With

• ai-security
• fastapi
• groq
• paypal
• python
• red-teaming
• security-testing
• universal-adapter