The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted text from audio transcripts, which creates a risk of indirect instructions influencing the agent. \n
Ingestion points: The agent reads transcript files from 1-原始文件/ and background documentation from 2-要求和信息/ as described in SKILL.md. \n
Boundary markers: The instructions lack explicit delimiters or directives to isolate ingested text from the agent's instructions. \n
Capability inventory: The skill uses read_file, write_file, shell execution via sed, and a Python script for its operations. \n
Sanitization: No sanitization or validation is performed on the content of the transcripts before processing. \n- [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands and a Python script with parameters that may be derived from user input. \n
The use of sed in the workflow (sed -n '开始行,结束行p' ...) involves direct interpolation of parameters into a shell string, which could be exploited for command injection if inputs are maliciously crafted. \n
The setup_workspace.py script accepts arbitrary file paths as arguments; if the agent is manipulated into passing sensitive system paths, the script will copy those files into the local workspace, potentially leading to unintended local data exposure.

long-audio-transcript-processor