The web_crawl tool performs its SSRF check only on the initial URL: it resolves the hostname once with socket.gethostbyname and rejects private/loopback/link-local results. It then passes the URL to a fetcher that uses httpx.Client(follow_redirects=True) - or urllib.request.urlopen when httpx is absent, which also follows redirects - and re-resolves the hostname at connect time, with no further validation. This validate-here/fetch-there gap is bypassable two independent ways: HTTP redirects and DNS rebinding.

Affected code: src/praisonai-agents/praisonaiagents/tools/web_crawl_tools.py

• Single-shot validation (lines 229-238): if os.environ.get("ALLOW_LOCAL_CRAWL") != "true": ip_str = socket.gethostbyname(hostname) # resolved ONCE, at validation time ip = ipaddress.ip_address(ip_str) if ip.is_loopback or ip.is_private or ip.is_link_local or ip.is_multicast or ip.is_unspecified: continue # rejected url_list.append(u)
• Vulnerable fetch (_crawl_with_httpx, lines 142 / 149): follows redirects, re-resolves DNS, no re-check: with httpx.Client(follow_redirects=True, timeout=30.0) as client: response = client.get(url)

  fallback: urllib.request.urlopen(url, timeout=30) (also follows redirects by default)

• web_crawl / crawl_web are registered tools (tools/init.py:156-157); httpx is the default fallback provider (dispatch at web_crawl_tools.py:269).

The two bypasses:

1. Redirect: validation approves an attacker domain resolving to a public IP; attacker server replies 302 Location: http://169.254.169.254/... (or any internal host); the fetcher follows it unchecked.
2. DNS rebinding (TOCTOU): validator's gethostbyname and fetcher's connect-time resolution are independent; a low-TTL attacker domain answers public to the validator and private/loopback to fetch.

Impact: An agent with web_crawl - driven by direct input or indirect prompt injection - can be made to read internal-only HTTP services and cloud instance-metadata endpoints (e.g. IAM credentials), with the response body returned in the tool output. Scope is Changed because the request pivots into the internal network.

Proof of concept: A PoC drives the real web_crawl() (httpx absent -> genuine urllib fallback). It runs a loopback "internal metadata" service and a loopback attacker redirector, substituting DNS only to stand in for "attacker owns a public domain" / offline routing - the redirect-following and connect-time re-resolution are the repo's own behavior. Observed: CONTROL: web_crawl("") -> blocked (validator works) PoC 1A (redirect): attacker.example approved (public); 302 -> loopback metadata -> result.content leaks {"AccessKeyId":"ASIA_FAKE_STOLEN_CREDENTIAL_..."} PoC 1B (rebinding): gethostbyname(rebind.example)->public (allowed); connect->127.0.0.1 -> same secret leaked The control proves the validator blocks a direct loopback request, so the bypasses are genuine.