Roshni Vijayakumar@Roshnivijay

TrustGuard AI is a security-focused agentic AI orchestration platform designed to protect multi-step AI agents from prompt injection and compromised tools. Modern AI agents can interact with websites, emails, databases, APIs, and other external tools. A malicious instruction hidden inside one of these sources can potentially influence the agent into performing actions unrelated to the user's original request. Our solution addresses this problem using information-flow control rather than relying on keyword-based prompt-injection detection. The user's request becomes the trusted source for generating the initial execution plan. Every value returned by a tool receives a provenance label identifying its origin and trust level. These labels persist when information is extracted, transformed, combined, or passed between tools. A policy enforcement layer evaluates whether data is permitted to reach sensitive tools such as internal-file access, email transmission, or financial operations. If an untrusted tool attempts to introduce an unrelated instruction or influence a sensitive action, the orchestrator blocks the operation and generates a security alert instead of allowing the agent to blindly follow the instruction. The prototype uses a scripted/LLM-driven agent with mocked tools to demonstrate both legitimate workflows and adversarial scenarios. We demonstrate at least five legitimate multi-tool tasks and five prompt-injection attacks, including reworded and multi-hop attacks where malicious instructions travel through multiple tools before reaching a sensitive operation. The system reports attack success rate, legitimate task completion rate, blocked actions, provenance information, and security decisions. This allows us to demonstrate that the security layer can contain malicious tool outputs while minimizing unnecessary blocking of legitimate agent activity. The key innovation is separating agent reasoning from security authorization. The LLM can propose an action, but the policy engine independently determines whether the requested information flow is permitted. This creates a practical trust boundary around the agent and its tools.

💻 Online Hackathon | ⏳ Build: 48 hours | 📅 September 25–27, 2026 The official IBM Bob 2.0 hackathon. Build what's next in AI-assisted development, alongside the developers and experts shaping it. ⭐ Get help from experts throughout the event 🤝 Join a team or build solo 💰 Prize pool: $12,000 🧑🏻‍💻 Register before the Kick-Off Stream to start building with Bob