PYSEC-2026-3396

Summary

The utcp-http plugin is vulnerable to a blind Server-Side Request Forgery (SSRF) caused by a trust-boundary inconsistency between manual discovery and tool invocation. register_manual() validates the discovery URL against an HTTPS / loopback allowlist, but call_tool() and call_tool_streaming() reuse the resolved tool_call_template.url directly without revalidating. An attacker who hosts a malicious OpenAPI spec on a legitimate HTTPS endpoint can declare servers: [{ url: "http://169.254.169.254" }] (or any internal address) in the spec; the OpenAPI converter blindly trusts that value and the tool becomes a blind SSRF primitive that exposes cloud metadata, internal services, and other firewalled-only endpoints to the LLM caller.

All three HTTP-class protocols (utcp_http.http, utcp_http.streamable_http, utcp_http.sse) shared the same gap, plus a separate prefix-bypass: the previous startswith("http://localhost") check let URLs like http://localhost.evil.com through.

Impact

A remote attacker who can convince the agent (via the LLM context, prompt injection, or a tool-discovery surface) to register their HTTPS OpenAPI URL can:

• Map internal networks behind the agent.
• Read AWS/GCP IAM credentials from cloud metadata endpoints (http://169.254.169.254, http://metadata.google.internal).
• Reach unauthenticated internal services (Elasticsearch, Redis HTTP, internal admin panels).
• Have responses returned to the LLM, which combined with prompt injection enables exfiltration back to the attacker.

Affected versions

utcp-http <= 1.1.1.

Patched versions

utcp-http 1.1.2.

Patch

Commit: 5b16e43 on dev.

• New utcp_http._security helper: ensure_secure_url(url, context=...) parses the URL with urllib.parse.urlparse and validates the hostname (not a string prefix) against the loopback set, closing the localhost.evil.com bypass.
• All three protocols call ensure_secure_url(url, context="manual discovery") in register_manual (replacing the duplicated prefix check) and ensure_secure_url(url, context="tool invocation") immediately before each aiohttp request in call_tool / call_tool_streaming. The runtime check is the actual SSRF fix.
• New regression tests in test_security.py pin the accept/reject decisions and explicitly cover the historical bypass cases.

Workarounds

For users who cannot upgrade immediately:

• Refuse to call register_manual with any URL controlled by an untrusted party, even over HTTPS.
• Restrict outbound network access from the host running the agent so internal addresses (RFC1918, 169.254.0.0/16, loopback for cloud metadata) are unreachable.

Credit

Discovered and reported by @YLChen-007 in #83.