brightdata-mcp
- Type
- MCP server
- Scanned
- Nov 12, 2026
- Version
- 4703ffa27d6b
- Files read
- 82
This MCP server connects an AI assistant to Bright Data's cloud services to search the web, scrape web pages and structured site data, and automate a remote browser. Its main risks are indirect prompt injection from untrusted web or profile data (notably search_dataset returning third-party LinkedIn text without trust-boundary markers) and runaway paid usage because the 14 remote browser tools bypass RATE_LIMIT.
What it can reach
- Secrets
- Reads the Bright Data API_TOKEN from environment variables and uses it to authenticate API calls and retrieve remote browser zone passwords.
- Network
- Sends HTTPS requests to api.brightdata.com and docs.brightdata.com to search engines, scrape public URLs, run pre-built scrapers, search LinkedIn datasets, query ChatGPT/Perplexity/Sophie, and create missing account zones on startup.
- Browser
- Connects over WebSocket CDP to Bright Data's remote Scraping Browser (brd.superproxy.io:9222) to navigate HTTP/HTTPS websites, click, fill forms, take screenshots, and read page content.
Findings
-
Mediumsearch_dataset returns untrusted third-party profile content without trust-boundary wrapping
All other external web, scraper, and browser tools add their names to external_content_tools so their text output is scrubbed of forged boundary markers and wrapped with a nonce-authenticated untrusted-data notice via wrap_untrusted(). Although comments in server.js state that dataset tools add themselves to external_content_tools, search_dataset is omitted. Any prompt-injection text placed by a stranger on a public LinkedIn person or company profile indexed in Bright Data's datasets is returned raw and unscrubbed to the AI agent.
Fix: Add external_content_tools.add('search_dataset') alongside the scraper tools in server.js so dataset search hits are scrubbed and wrapped as untrusted content.
-
LowRemote browser automation tools bypass RATE_LIMIT enforcement and session_stats tracking
Every non-browser tool is registered through tool_fn(), which calls check_rate_limit() to enforce the user's RATE_LIMIT setting and records call counts in debug_stats for session_stats. The 14 scraping_browser_* tools use a separate inline execute wrapper that omits check_rate_limit() and debug_stats updates. If a user configures RATE_LIMIT to cap requests or spending, a prompt-injected or looping agent can still make unlimited paid remote browser calls (billed per GB), and session_stats will show zero browser calls.
Fix: Register browser_tools through tool_fn(tool.name, tool.execute, BROWSER_RULES) or call check_rate_limit() and update debug_stats inside the browser tool wrapper.
-
LowMCP configuration content gate skips pre-built scraper and dataset tools
By default, CONTENT_GATE inspects returned web content for instructions trying to trick the coding agent into editing MCP config files or running npx -y packages, pausing to ask the user via MCP elicitation. However, page_content_tools only covers scrape_as_markdown, search_engine, and their batch variants, while the server's own instructions direct the agent to prefer run_scraper (and get_scraper_results) for 1,200+ sites including GitHub repositories, npm, PyPI, and Reddit. Malicious MCP config instructions fetched via run_scraper, get_scraper_results, or search_dataset bypass this gate without prompting the user.
Fix: Extend gate_page_content() to cover run_scraper, get_scraper_results, and search_dataset, and explicitly document any remaining excluded tools in README.md.
-
LowExternal API and scraped-data tools advertise openWorldHint as false
Tools that send data to external Bright Data endpoints or return scraped web content—specifically get_scraper_results, search_scrapers, get_scraper_details, get_scraper_progress, refresh_scrapers, search_dataset, and ask_brightdata_assistant (which sends up to 20 KB of question text to an external assistant API)—set readOnlyHint: true and openWorldHint: false, the same hints used for local in-memory session_stats. In MCP clients that auto-approve closed-world read-only tools while prompting on open-world network tools, a prompt-injected agent can invoke these external network calls without user confirmation.
Fix: Set openWorldHint: true on tools that communicate over the network with external APIs or retrieve third-party web records, especially get_scraper_results, search_dataset, and ask_brightdata_assistant.
-
LowBrowser session error logging prints unredacted error objects to stderr
While connect_over_cdp() and server.js redact wss:// customer and password credentials from error messages returned to the agent, catch blocks in Browser_session.get_browser(), get_page(), and close() pass the raw Error object directly to console.error() on stderr. If Playwright throws an error after initial connection (such as a dropped CDP WebSocket during newPage() or close()) whose message or stack includes the CDP endpoint URL, the browser zone password can be written unredacted to local MCP logs.
Fix: Pass error messages and stacks through redact_url_creds() before logging them with console.error() in browser_session.js.
-
LowRelease workflow downloads unpinned mcp-publisher binary without checksum verification
.github/workflows/publish-mcp.yml:34
The GitHub Actions workflow that publishes metadata to the MCP Registry downloads the latest mcp-publisher tarball over curl and executes it without pinning a release tag or verifying a SHA-256 digest, inside a job granted id-token: write permissions. Installed packages never run this workflow, so it only affects maintainer release infrastructure.
Fix: Pin mcp-publisher to a specific release version and verify its SHA-256 checksum before extracting and executing the binary.
-
InfoServer automatically creates SERP and Browser zones on the Bright Data account at startup
Every time server.js starts, ensure_required_zones() checks the user's Bright Data account and sends POST requests to create the mcp_serp and mcp_browser zones if they do not exist, even when GROUPS or TOOLS has disabled search or browser tools. This account modification is not mentioned in README.md.
Fix: Only create zones required by the enabled tool set (or lazily when a search or browser tool is called) and document automatic zone creation in README.md.
