just-scrape
ScrapeGraphAI/just-scrape/skills/just-scrape
- Type
- Skill
- Scanned
- Nov 13, 2026
- Version
- a3d179ffe269
- Files read
- 1
This skill teaches an AI coding agent to search, scrape, crawl, extract structured data from, and monitor websites through the ScrapeGraph AI CLI (just-scrape). Its biggest risk is that it grants unrestricted Bash execution while reading untrusted web pages, so prompt injection on a scraped site could trick the agent into running arbitrary shell commands without asking.
What it can reach
- Shell commands
- Grants unrestricted Bash access while the skill is active to run just-scrape, npm, mkdir, jq, rg, head, wc, and shell redirection.
- Network
- Sends search, scrape, crawl, extract, and recurring monitor requests (including optional webhook URLs) to ScrapeGraph AI's cloud API (https://v2-api.scrapegraphai.com or SGAI_API_URL).
- Files
- Creates a .just-scrape directory in the current workspace to save and read JSON output, suggests editing .gitignore, and relies on CLI configuration in .env or ~/.scrapegraphai/config.json.
- Secrets
- Uses SGAI_API_KEY to authenticate with ScrapeGraph AI and instructs the agent to read environment variables such as API_TOKEN and SESSION_COOKIE for authenticated page requests.
Findings
-
HighUnrestricted Bash auto-approval while handling untrusted web content
The skill frontmatter requests allowed-tools: Bash without restricting commands to the just-scrape binary. In coding agents that honor skill tool allowlists (such as Claude Code), this lets the agent run any shell command without asking for confirmation while the skill is active. Because the skill fetches and reads untrusted third-party web pages, hidden prompt-injection text on a scraped page could trick the agent into running harmful commands despite the warning in the Security section.
Fix: Restrict allowed-tools to Bash(just-scrape:*) and use built-in read/search tools instead of arbitrary shell commands so non-CLI commands still require user approval.
-
MediumAutomatic runtime global install of an unpinned npm package
skills/just-scrape/SKILL.md:32
When just-scrape is not already installed, the Prerequisites and Troubleshooting sections instruct the agent to automatically install and run just-scrape@latest from npm. Because @latest is not pinned to a reviewed version and npm runs package install scripts by default, a compromised update to just-scrape or one of its dependencies would run arbitrary code on the user's machine at runtime.
Fix: Remove automatic npm install -g just-scrape@latest and npx just-scrape@latest instructions, and ask the user to manually install a pinned version before running the skill.
-
MediumSession cookies and bearer tokens exposed in process arguments and sent to third-party servers
skills/just-scrape/SKILL.md:128
The skill instructs the agent to scrape authenticated pages by passing double-quoted environment variables such as $SESSION_COOKIE and $API_TOKEN to --cookies and --headers. First, the shell expands these variables before starting the CLI, exposing plaintext secrets in the local process list (ps) to other users or monitoring tools on the machine. Second, just-scrape does not fetch pages locally; it sends requests through ScrapeGraph AI's cloud servers (where request payloads can also be retained in history), forwarding the user's third-party session cookies and bearer tokens to an external service without warning.
Fix: Warn users clearly that --cookies and --headers transmit credentials to ScrapeGraph AI's cloud servers, and avoid passing expanded secrets directly on the command line.
