The Agent Skills Directory

[DYNAMIC_EXECUTION]: The diagnostic scripts scripts/incident_artifact_tool.py and scripts/replay_trusted_request_dump.py use the pickle.load() function to process request and crash dump files provided via command-line arguments. Python's pickle module is inherently unsafe and allows for the execution of arbitrary code during deserialization. The script scripts/replay_trusted_request_dump.py explicitly states that it uses plain pickle to bypass security restrictions found in newer SGLang builds, which increases the risk of execution from untrusted dump files.
[REMOTE_CODE_EXECUTION]: The insecure use of pickle.load() provides a direct path for remote code execution. If the agent is instructed to summarize or replay a maliciously crafted dump file provided by an attacker, it will execute the embedded code on the host system.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from live SGLang server endpoints (such as /metrics and /v1/loads) using scripts/incident_artifact_tool.py. It lacks boundary markers or sanitization when summarizing this external data, which could allow malicious content within server responses to influence the agent's reasoning or actions. The skill's capabilities include local filesystem access, network requests, and the execution of the included Python scripts.

sglang-prod-incident-triage