Executive brief
PraisonAI's web_crawl tool—a standard feature used by AI agents to fetch web content—only validates the initial URL's IP address against a blocklist of private and internal networks. However, when the server responds with an HTTP redirect, the tool follows it without re-validating the destination, allowing attackers to redirect legitimate public URLs to cloud metadata services or internal systems. An attacker who can influence what the agent crawls (through task instructions or prompt injection) can steal cloud IAM credentials, access localhost services, and retrieve other internal data.
Technical details
The vulnerability is a classic SSRF (CWE-918) bypass in the web_crawl tool's HTTP redirect handling. The tool validates only the initial hostname against a blocklist of loopback, private, link-local, multicast, and unspecified IP ranges using socket.gethostbyname() at line 231 of web_crawl_tools.py. However, the HTTP client (httpx.Client with follow_redirects=True at line 142) automatically follows 3xx redirects without re-validating the redirect target against the same blocklist. An attacker supplies a public URL that responds with a 302 Location header pointing to an internal address (e.g., 169.254.169.254 for AWS metadata or 127.0.0.1 for localhost), bypassing the initial IP check. The fetched internal response body is returned into the agent context and exposed to the model, logs, and downstream tools. This is the default behavior with httpx as the crawl provider; the vulnerability requires only that ALLOW_LOCAL_CRAWL is not explicitly set to true and the agent's crawl target is controllable by an attacker.
Affected products
- MervinPraison praisonaiagents < 1.6.58
Timeline
- 2026-08-25: disclosed
- 2026-08-25: patched: Version 1.6.58 patches the vulnerability
- 2026-08-25: kev added