Latest Ollama vulnerability: CVE-2026-7482 #

Researchers disclosed that certain versions of Ollama are susceptible to a heap out-of-bounds read vulnerability within the GGUF model loader. A remote, unauthenticated attacker could exploit this by sending a specially crafted GGUF file to the /api/create endpoint. When the server processes a GGUF file where the declared tensor offset and size exceed the file's actual length, the functions in fs/ggml/gguf.go and server/quantization.go (WriteTo()) read past the allocated heap buffer during the quantization process.

The resulting memory leak may expose sensitive information, including environment variables, API keys, system prompts, and concurrent user conversation data. This data can then be exfiltrated by uploading the resulting model artifact to an attacker-controlled registry via the /api/push endpoint. In the upstream distribution, the /api/create and /api/push endpoints lack authentication. While default deployments bind to 127.0.0.1, the documented OLLAMA_HOST=0.0.0.0 configuration is common in practice, leading to significant public Internet exposure. This vulnerability has been designated CVE-2026-7482, also known as Bleeding Llama, and has been rated critical with a CVSS score of 9.1.

    The following versions are affected:

    • Ollama: Versions prior to 0.17.1

      What is Ollama? #

      Ollama is an open-source framework designed for the local deployment, management, and execution of large language models (LLMs) on personal computing hardware.

      What is the impact? #

      Successful exploitation of this vulnerability could expose sensitive information, including environment variables, API keys, system prompts, and concurrent user conversation data.

      Are updates or workarounds available? #

      Users are encouraged upgrade affected systems to the following versions immediately:

      • Ollama: Upgrade to 0.17.1 or later.

      How to find potentially vulnerable systems with runZero #

      From the Software Inventory, use the following query to locate potentially impacted assets:

      vendor:=Ollama AND product:=Ollama AND source:runzero

      Written by Matthew Kienow

      Matthew Kienow is a software engineer and security researcher. Matthew previously worked on the Recog recognition framework, AttackerKB as well as Metasploit's MSF 5 APIs. He has also designed, built, and successfully deployed many secure software solutions; however, often he enjoys breaking them instead. He has presented his research at various security conferences including DerbyCon, Hack In Paris, and CarolinaCon. His research has been cited by CSO, Threatpost and SC Magazine.

      More about Matthew Kienow
      Subscribe Now

      Get the latest news and expert insights delivered in your inbox.

      Welcome to the club! Your subscription to our newsletter is successful.

      Explore more runZero

      Product
      Announcing runZero 5.0: Exposure management built to outpace AI-driven attacks
      When you're up against AI, every minute counts. Get deep, actionable intelligence across your entire attack surface to close the gaps and hold the...
      Product Videos
      runZero 5.0: Platform Demo
      With the new 5.0 release, runZero is giving defenders the edge they need to succeed in the AI-attack era.
      runZero Perspective
      BOD 26-04: A new era of prioritized remediation
      A complete breakdown of CISA's BOD 26-04 directive. Learn how the shift to SSVC, risk-based KEV prioritization, and 3-day remediation impacts your...
      runZero Perspective
      Dawn of the apex agentic adversary
      When agentic AI can weaponize exploits in seconds, visibility is everything. Stop the predator with runZero’s exposure management for the AI-attack...
      Webcasts
      runZero Hour, Ep. 32: AI-pocalypse now? Why the 2026 DBIR is actually good news
      In this episode of runZero Hour, Tod Beardsley, Brianna Cluck, and Verizon's Alex Pinto broke down 2026 DBIR trends, AI threats, and runZero 5.0.
      Podcasts
      The shadow era AI, exploits, and cybersecurity's 90s comeback
      HD Moore explains how AI is turning hacking back to the 90s, generating permanent exploit skeleton keys and breaking traditional defense.
      Talks
      Identifying exposures at scale with BloodHound OpenGraph
      Traditional exposure management misses hidden attack paths. Learn how BloodHound and Cypher queries can uncover vulnerabilities beyond individual...
      Podcasts
      When is a vulnerability not a vulnerability?
      Attackers care about outcomes, not CVEs. Tod Beardsley joins Distilled Security to discuss reframing risk from checklist compliance to adversary...

      See Results in Minutes

      See & secure your total attack surface. Even the unknowns & unmanageable.