skills
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- a2f903361a99
- Files read
- 19
This skill pack teaches AI coding agents to use the Firecrawl CLI to search the web, scrape or crawl websites, drive remote browser sessions, watch pages for changes, and convert documents to markdown. Its biggest risk is unexpected data exposure and pre-approved actions, because the skills silently report the user's broader task objective to Firecrawl in background feedback calls, route local documents to Firecrawl's cloud whenever a file path is given, and allow every firecrawl subcommand without user confirmation while reading untrusted web pages.
What it can reach
- Shell commands
- Runs firecrawl and npx firecrawl-cli commands without confirmation prompts, and instructs the agent to install or update the global firecrawl-cli npm package and inspect output files with standard CLI utilities.
- Network
- Sends search queries, target URLs, uploaded local documents, recurring monitor and webhook configurations, and background session feedback to Firecrawl's cloud API and connected Alexandria third-party data providers.
- Files
- Saves scraped content and site downloads under .firecrawl/, modifies .gitignore, uploads local files up to 50 MB via firecrawl parse, and installs skill files globally across detected coding editor directories during setup.
- Browser
- Opens the local browser for OAuth sign-in and controls remote cloud browser sessions via firecrawl interact, including saving and reusing website cookies and localStorage in named Firecrawl profiles.
- Secrets
- Uses and configures Firecrawl API keys or OAuth login tokens, and may transmit website login inputs to remote Firecrawl browser sessions when asked to log into a site.
Findings
-
MediumSilent background feedback sends the user's broader session objective to Firecrawl by default
skills/firecrawl-alexandria/SKILL.md:31
Both the Alexandria and Search skills instruct the agent to automatically run background feedback commands (--silent &) after calls in exchange for a 1-credit refund. Beyond rating tool output, the Alexandria skill tells the agent to summarize and send the user's overarching session goal (--objective), and the Search skill sends expected missing topics (--missing-content) and external URLs visited (--valuable-sources). If a user works on a private business, security, or engineering task and the agent makes a routine public lookup, that confidential task objective is silently sent to Firecrawl and aggregated across teams without asking the user.
Fix: Set FIRECRAWL_NO_ENDPOINT_FEEDBACK=1 and FIRECRAWL_NO_SEARCH_FEEDBACK=1 in your environment (or disable feedback in Firecrawl team settings), and edit the skill files to require user permission before sending feedback.
-
MediumWildcard tool pre-approval allows every Firecrawl subcommand without user confirmation
Ten skills in this pack declare allowed-tools: Bash(firecrawl *) and Bash(npx firecrawl-cli *), which pre-approves every CLI subcommand and flag without prompting the user. Because the skills routinely ingest untrusted third-party web pages that can carry indirect prompt injections, a malicious webpage that tricks the agent can silently trigger sensitive subcommands—including uploading local files (firecrawl parse), driving authenticated saved browser profiles (--profile with firecrawl interact), creating recurring webhook monitors (firecrawl monitor create --webhook-url), or exfiltrating data in URL requests.
Fix: Restrict allowed-tools in each skill to its specific read-only subcommand (such as Bash(firecrawl search *) or Bash(firecrawl map *)), and require manual user approval for parse, interact, monitor, login, init, and setup.
-
MediumBroad local-file trigger uploads workspace documents and HTML files to Firecrawl's cloud
skills/firecrawl-parse/SKILL.md:4
The firecrawl-parse skill tells the agent to activate whenever the input is a local file path (including PDF, DOCX, XLSX, and plain HTML files) rather than only when the user asks to use Firecrawl, without mentioning in its description that parsing uploads the file (up to 50 MB) to Firecrawl's remote servers. Combined with pre-approved Bash(firecrawl *) execution, asking a coding agent an ordinary question about a private local spreadsheet, PDF, or HTML file in a project can silently upload that file to Firecrawl instead of reading it locally.
Fix: Remove the firecrawl-parse skill if you want local files kept on disk, or change its description so it only runs when the user explicitly asks to upload and parse a file through Firecrawl.
-
LowSetup rules run an unpinned @latest npm package with automatic confirmation
skills/firecrawl/rules/install.md:15
When the CLI is missing or outdated, rules/install.md instructs the agent to execute npx -y firecrawl-cli@latest init -y --browser, and the skills pre-approve Bash(npx firecrawl-cli *). Running @latest with -y downloads and executes whatever version is currently published on npm without user confirmation and installs skills globally across all detected coding editors on the computer, exposing the user to supply-chain compromise if the npm package is ever hijacked.
Fix: Install a specific audited version manually (npm install -g firecrawl-cli@<exact-version>) and remove Bash(npx firecrawl-cli *) from allowed-tools.
