firecrawl-research-patterns

Pass

Audited by Gen Agent Trust Hub on Nov 7, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPERSISTENCEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves scraping arbitrary web pages and academic papers, which are then processed by an LLM to extract findings. This creates a large surface for indirect prompt injection from untrusted external sources.\n
  • Ingestion points: Web content retrieved via the Firecrawl public API, Jina Reader, and academic sites like arXiv, IEEE, and ACM.\n
  • Boundary markers: The skill lacks instructions for using delimiters or boundary markers when passing scraped content to the LLM for summarization or synthesis, increasing the likelihood that the agent might follow instructions embedded in the source material.\n
  • Capability inventory: The agent possesses Write, Edit, and Bash capabilities, allowing it to modify files or execute system commands if manipulated by injected instructions.\n
  • Sanitization: The instructions do not define any sanitization or filtering steps for the scraped markdown content before it is processed by the agent.\n- [PERSISTENCE]: The skill includes a 'Self-Evolving Skill' directive that instructs the agent to 'fix this file immediately' if errors are found or parameters change. This enables the agent to autonomously modify its own persistent instruction set.\n- [PROMPT_INJECTION]: The 'Self-Evolving' mechanism acts as a meta-instruction that could allow malicious input from the web to influence the agent to permanently alter its behavior by rewriting the skill's markdown file.\n- [COMMAND_EXECUTION]: The skill provides numerous shell command templates using curl, jq, grep, tar, and pandoc for API interaction, data processing, and archive management.\n- [EXTERNAL_DOWNLOADS]: The skill utilizes several external services for data acquisition, including api.firecrawl.dev, api.semanticscholar.org, r.jina.ai, and arxiv.org. It also references the p-limit Node.js package and the pandoc CLI tool.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 7, 2026, 02:13 AM