The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md (Step 5) construct a shell command by interpolating user-provided company names and technical fields directly: python "<SKILL_DIR>/scripts/sogou_search.py" "{公司名} {技术领域}" 8. This pattern is susceptible to command injection if the input contains shell metacharacters such as semicolons or pipes.
[DYNAMIC_EXECUTION]: The skill implements a workflow where the agent writes a Python file (v2_data.py) containing retrieved external data and then runs a separate script (build_report_v2.py) which imports this file. This mechanism effectively executes data as code, presenting a risk of arbitrary code execution if malicious payloads are ingested from external sources such as patent titles, news summaries, or academic literature.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources including patent databases, academic literature, and web search results. This untrusted data is summarized by the agent and incorporated into reports.
Ingestion points: Patent details from mcp__zhihuiya__search_patent, literature from mcp__zhihuiya__search_literature, and news results from sogou_search.py.
Capability inventory: Shell command execution via python calls and file writing to the user's document directory.
Boundary markers: No explicit delimiters or instructions are used to separate untrusted external content from the agent's logic.
Sanitization: Although the final HTML report uses escaping, the intermediate Python code generation step lacks safeguards against code injection.
[EXTERNAL_DOWNLOADS]: The scripts/sogou_search.py script performs network requests to sogou.com to retrieve industry news data.

tech-briefing