competition-prompt-injection

To trace prompt injection paths, map boundaries from system prompts and user input through retrieval, memory, planner, and tool adapters to identify where unsafe transitions occur. This boundary mapping captures compact evidence blocks detailing malicious chunks and final tool invocations.

Planner drift occurs when an agent's task planning deviates due to malicious content. You analyze planner drift by tracing untrusted content visible to the model as it travels through planner and tool interfaces, capturing evidence of the drift and resulting tool-argument corruption.

Detect retrieval poisoning and memory contamination by identifying untrusted content and tracing its path through retrieval and memory interfaces. This analysis spots boundary violations and captures evidence detailing the malicious chunk causing the contamination.

Yes, this analysis applies after a sandbox orchestrator has established sandbox assumptions and routing. It requires an active sandboxed agent chain to provide traceable evidence and sandbox-aware guidance for mapping boundaries across system, memory, and tool interfaces.

Investigate tool-boundary abuse and agent exfiltration by tracing unsafe transitions from system prompts through tool adapters. This captures evidence of final tool invocations and maps boundaries across tool interfaces to determine actionable mitigations for secret exposure.

The best way to analyze tool-argument corruption is identifying the first untrusted content visible to the model and tracing its path through the system. This evidence-driven approach captures malicious chunks, planner drift, and final tool invocations to spot boundary violations.