llama.cpp v1.0.0 hardens llama-server against SSRF and unauthenticated tool access

llama.cpp v1.0.0 refuses to start llama-server with tools or MCP on a non-loopback address unless an API key is set, and stops the CORS proxy and media URLs reaching private addresses.

Christopher Ort
By Christopher Ort · 2 min read
Share
A developer working at a laptop in a home office, with network cables and a router visible on the desk
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. Who is affected
  5. What you should do
  6. Sources
i10X Quick Take

Quick Take

  1. What: llama.cpp v1.0.0 blocks server-side request forgery through the CORS proxy and media URLs, and requires API keys for tools or MCP on non-loopback addresses
  2. Who is affected: llama-server builds before v1.0.0, especially machines that also browse the web or servers bound to 0.0.0.0 with tools enabled and no API key
  3. What to do: Update to llama.cpp v1.0.0; set an API key if tools or MCP are enabled on a non-loopback address; enable CORS on local MCP servers if the CORS proxy was used to reach them
  4. Fixed in: v1.0.0, released November 13, 2026
  5. Found by: Xuan-Son Nguyen, llama.cpp maintainer

In plain English

The llama.cpp project released version 1.0.0 with security fixes for its llama-server component. Before this version, a visited web page could use the server’s CORS proxy to reach internal network addresses, which is called server-side request forgery. The new version also refuses to start the server with tools or MCP enabled on a network address unless you set an API key.

What you should do: If you run llama-server on your personal computer or on a network, you should update to v1.0.0. If you don’t use llama.cpp or llama-server, you can ignore this.

What happened

The llama.cpp project released v1.0.0 on November 13, 2026 with a group of server security changes. Maintainer Xuan-Son Nguyen wrote and merged the two main pull requests, 32113 and 32115, on November 12, 2026. Neither change carries a CVE.

The CORS proxy and the media URLs accepted in chat completion input could previously resolve to any address, including loopback and LAN hosts, which is server-side request forgery. Version 1.0.0 refuses private targets unless the operator passes --media-allow-private-urls. A reviewer reported that both a redirect to 127.0.0.1 and a DNS rebinding host were blocked.

Separately, starting the server with tools or MCP enabled on a non-loopback address and no API key is now refused. CORS origin matching now checks the port as well as the host.

Nguyen writes that the original design assumed operators would sandbox llama-server, while many now run it on personal computers.

Who is affected

Builds before v1.0.0 are exposed on any machine that also browses the web, because a visited page can drive the CORS proxy at internal addresses. Anyone who bound the server to 0.0.0.0 with tools enabled and no key was serving those tools to the network.

What you should do

Update to llama.cpp v1.0.0. If start-up now fails because tools or MCP are enabled on a non-loopback address, set an API key rather than passing --allow-unauthenticated-tools. Anything that relied on the CORS proxy to reach a localhost or LAN MCP server needs CORS enabled on that server instead.

Sources

Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. GAO Report: ChatGPT, Claude, Gemini and Grok Produced Viable Attack Plans After Safeguards Were Bypassed
  2. ResolveAI SRE-1: A 27B Model Trained to Find the Root Cause of Production Incidents
  3. Reson8 Rumble-2-Flash: Open-Weight 0.5B Speech Recognition Model Covers 40+ Languages
  4. Pika Enact Lets Creators Direct AI Video Actors and Cameras in Real Time
  5. Google Stitch Loop: An Always-On Product Agent That Fixes Security, Speed and Accessibility via Pull Requests