The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts provided with the skill utilize the Python pickle module to load data from external files without safety restrictions, enabling potential arbitrary code execution.
Evidence: The load_dump_payload function in scripts/incident_artifact_tool.py and the load_records function in scripts/replay_trusted_request_dump.py both call pickle.load() on user-supplied files.
The script scripts/replay_trusted_request_dump.py explicitly states it uses "plain pickle loading to bypass SafeUnpickler restrictions."
[COMMAND_EXECUTION]: The skill requires the agent to execute various shell commands and scripts to perform its intended triage functions.
Evidence: SKILL.md provides multiple executable blocks involving python3 scripts/incident_artifact_tool.py, cuda-gdb, and git bisect.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external API endpoints and incident bundles, creating a significant attack surface for indirect injection.
Ingestion points: API response bodies from the base_url (e.g., /health, /metrics, /v1/loads) processed in scripts/incident_artifact_tool.py.
Boundary markers: None explicitly defined to prevent the agent from obeying instructions embedded in the logs or metrics.
Capability inventory: The skill has extensive capabilities including file system writes to /tmp, network requests, and shell command execution.
Sanitization: No sanitization or integrity verification is performed on the content of the bundles or dumps before they are summarized or replayed.

sglang-prod-incident-triage