← Back to overview

CVE-2026-77775 affects Headroom's LLM proxy, which allows clients to specify an arbitrary upstream destination via the x-headroom-base-url request header. The proxy functions _resolve_openai_upstream_base and _select_passthrough_base_url accept this header without validating the destination, permitting requests to loopback, link-local, and RFC 1918 (private) addresses, enabling Server-Side Request Forgery (SSRF). Internal services and cloud metadata endpoints (e.g., AWS IMDSv1) can be reached and their responses returned to the caller. Additionally, the Authorization header is forwarded unchanged to the attacker-designated host, risking credential leakage. The default pip install binds to 127.0.0.1, but the shipped docker-compose.yml exposes the service on 0.0.0.0 without requiring HEADROOM_PROXY_TOKEN authentication, greatly widening the attack surface. A fix was introduced in v0.36.1 via upstream_guard.py. Deployments following the reference compose configuration are exposed to unauthenticated network-level exploitation.

Affected products

  • Headroom LLM Proxy

Related CVE's

  • CVE-2026-77775

Categories

  • Cloud & Virtualization
  • Emerging Technologies
  • Identity & Access
  • Web Technologies