# David Castro - Agent Guard

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [David Castro](https://cerebralvalley.ai/u/David506)
- **GitHub:** https://github.com/dceciliano/agentguard
- **Demo video:** https://youtu.be/0xImDbFTHBQ
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4408

AgentGuard — Security Testing for AI Agents

AI agents are increasingly connected to business applications, sensitive information, and tools capable of performing real-world actions. However, these capabilities introduce a critical security challenge: untrusted content can contain malicious instructions that attempt to manipulate agent behavior and trigger unauthorized operations.

AgentGuard is an AI agent security testing platform designed to help developers identify, investigate, and mitigate these risks before deployment.

The platform evaluates agent behavior against indirect prompt-injection scenarios, validates tool execution against authorization boundaries, and produces explainable security findings supported by execution traces.

Its MVP includes:

Security Assessments: Evaluate hardened and intentionally vulnerable agent configurations.
Tool Authorization Analysis: Distinguish legitimate task authorization from technical tool-execution permissions.
Evidence-Based Findings: Detect unauthorized actions and report their severity, execution status, and supporting trace evidence.
Security Comparisons: Compare configurations to identify observed improvements in security controls.
Interactive Dashboard: Review assessments, findings, execution timelines, and mitigation results.
Meta Muse Spark Integration: Support live model-driven testing through the Meta Muse API.

The demonstration uses deterministic offline fixtures to reproduce a prompt-injection scenario in which an agent workflow attempts an unauthorized mock refund. AgentGuard detects the unauthorized action and compares the vulnerable outcome with a hardened configuration.

For transparency, the demonstrated unauthorized refund is simulated and does not represent a successful compromise of Meta Muse Spark.

AgentGuard's long-term vision is to become a continuous security validation platform for AI agents, enabling developers and security teams to test agent workflows, validate policy enforcement, and build safer AI-powered applications.

## More from Meta Global AI Developer Hackathon

- [Innovian](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4405)
- [Stardive](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4406)
- [Ishan's Team](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4407)
- [dark%hunter](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4409)
- [Alone](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4410)
- [Sijil](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4411)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4408. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt