The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from torch profiler traces and external JSON configuration files to generate performance reports. Maliciously crafted data within these traces, such as deceptive function names or metadata, could potentially influence the agent's interpretation of the results. 1. Ingestion points: The scripts extract_compute_flow_from_trace.py and model_compute_simulator.py read user-provided files via command-line arguments like --input or --kernel-flow. 2. Boundary markers: The output generated for the agent does not use specific delimiters or 'ignore instructions' warnings to prevent the agent from following directions potentially embedded in the trace data. 3. Capability inventory: The skill's functionality is limited to compute simulation and data processing; it does not request network access or have the ability to modify sensitive system files. 4. Sanitization: The skill extracts and displays metadata from the trace files without filtering for potential prompt injection patterns.

model-compute-simulation