惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

J
Java Code Geeks
量子位
MongoDB | Blog
MongoDB | Blog
N
Netflix TechBlog - Medium
奇客Solidot–传递最新科技情报
奇客Solidot–传递最新科技情报
B
Blog
A
About on SuperTechFans
腾讯CDC
The GitHub Blog
The GitHub Blog
云风的 BLOG
云风的 BLOG
雷峰网
雷峰网
Last Week in AI
Last Week in AI
H
Help Net Security
WordPress大学
WordPress大学
博客园 - 司徒正美
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
H
Hackread – Cybersecurity News, Data Breaches, AI and More
T
Tailwind CSS Blog
博客园 - 【当耐特】
S
SegmentFault 最新的问题
美团技术团队
M
MIT News - Artificial intelligence
L
LangChain Blog
博客园 - 聂微东

Help Net Security

Police arrest 10 suspected members of Black Axe cybercrime gang ShinyHunters claims it stole 1.4 million records from Udemy Sevii unveils Cyber Swarm Defense Mode to stop AI-driven attacks at scale Alleged Chinese hacker extradited to US over cyberattacks targeting COVID-19 research Cequence Agent Personas bring granular control and governance to enterprise AI agents NowSecure MARI gives enterprises evidence-based visibility into third-party mobile app risk The metrics killing your SOC, and what to use instead US state privacy fines reached $3.425 billion in 2025 Canada’s first SMS blaster case leads to three arrests Linux storage management tool Stratis 3.9.0 adds online encryption and cache-less pool startup TLS Connect gives SMBs a right-sized automated tool to manage TLS certificates Aptori expands its platform with autonomous offensive testing to reduce security bottlenecks Your IAM was built for humans, AI agents don’t care The AI criminal mastermind is already hiring on gig platforms 25 open-source cybersecurity tools that don’t care about your budget Product showcase: LuLu reveals unauthorized outbound connections from Mac apps Week in review: Claude Mythos finds 271 Firefox flaws, Vercel breach Users advised to drop passwords and make room for passkeys - Help Net Security Indirect prompt injection is taking hold in the wild - Help Net Security Compromised everyday devices power Chinese cyber espionage operations - Help Net Security New Cisco firewall malware can only be killed by pulling the plug - Help Net Security Meta is overhauling how you sign in, manage settings, and protect your accounts - Help Net Security Ubuntu 26.04 LTS delivers memory-safe system tools and live patching for Arm servers - Help Net Security OpenAI’s GPT-5.5 is out with expanded cybersecurity safeguards - Help Net Security AI is speeding up nation-state cyber programs - Help Net Security A study of 1,000 Android apps finds a privacy policy logging gap - Help Net Security IT spending to hit $6.31 trillion record, thanks to AI - Help Net Security Where AI in CI/CD is working for engineering teams - Help Net Security With AI's help, North Korean hackers stumbled into a near-undetectable attack - Help Net Security Hacker with a special interest in breaching sports institutions ends behind bars - Help Net Security
Unpatched flaws turn Ollama’s auto-updater into a persist...
Zeljka Zorz · 2026-05-05 · via Help Net Security

Researchers at Striga have disclosed two vulnerabilities (CVE-2026-42248, CVE-2026-42249) in Ollama’s Windows auto-updater that, when chained together, may allow an attacker to covertly plant a persistent executable that runs on every login.

Ollama Windows CVE-2026-42248 CVE-2026-42249

CVE-2026-42248 and CVE-2026-42249

Ollama is an open-source tool for running large language models locally. It’s is used by those who don’t want their data to leave their machine and don’t want to be constrained by API costs, usage limits, or the requirement of an internet connection.

“The Windows desktop client bundles the local Ollama server, a chat UI, and a system-tray icon. It auto-starts on login and checks for updates in the background. Striga surfaced the [vulnerability] chain during an audit of the Ollama repository,” Striga co-founder Bartłomiej Dmitruk explained.

CVE-2026-42248 is straightforward: the Windows build’s auto-updater signature verification function exists, it gets called, but it does nothing, and whatever is downloaded gets executed. (The macOS Ollama build performs proper code-signing checks, the researchers found.)

CVE-2026-42249 is a path traversal flaw, and stems from the fact that Ollama’s Windows updater builds the local path for a staged installer directly from HTTP response headers without sanitizing them.

An attacker who controls the update response can supply a malicious ETag header with ../ sequences and write an arbitrary executable straight into the user’s Windows Startup folder, Dmitruk noted.

Because the signature check returns “no error”, the post-write cleanup that would normally remove an unsigned file never runs, and the payload survives indefinitely. On the next login, Windows runs whatever was dropped there, and there is no warning (since the dropped file does not carry a Mark-of-the-Web tag).

“The same dropped binary fires on every subsequent login until the file is removed,” he pointed out, and added that the missing verification without the path traversal also provides remote code execution through Ollama’s normal update flow, but no persistence.

Attack prerequisites

For successful exploitation, an attacker needs to control the update response Ollama receives.

Dmitruk identified three paths: by compromising the update infrastructure directly; redirecting the client to an attacker-controlled server via a local foothold (the OLLAMA_UPDATE_URL variable, a hosts file edit, or a rogue root certificate); or network-level interception via TLS MITM or DNS hijacking with a forged certificate.

Other attack requirements include the Auto-update option being on (it’s on by default) and Ollama “sitting” in the Startup folder (it is by default).

“For the secondary path (CVE-2026-42248 alone, without traversal), Ollama additionally needs to be launched from the Startup-folder shortcut so DoUpgradeAtStartup fires. The primary chain does not need this; Windows runs the dropped file from the user’s Startup folder regardless of Ollama,” he told Help Net Security.

Striga researchers reported the findings to Ollama’s documented security address in late January 2026, Dmitruk said, and that address never replied. A maintainer’s personal email yielded one acknowledgement, then silence, he told us. “We have not seen any public statement from the maintainers about either CVE.”

After five weeks with no engagement, CERT Polska took over coordination of disclosure, assigned the CVEs, and published a warning on April 29, confirming that Ollama for Windows versions 0.12.10 through 0.17.5 are vulnerable.

What should users do?

“CERT Polska tested 0.15.1 end-to-end with our PoC and listed that as confirmed-vulnerable, deliberately cautious about untested releases,” Dmitruk told Help Net Security.

“We extended the verification: ran static checks across every release tag for the four chain indicators – (filepath.Join(UpdateStageDir, etag, …), verifyDownload(){ return nil }, OLLAMA_UPDATE_URL override, STARTF_TITLEISLINKNAME detection) – and confirmed all four present from v0.12.10 (when the rewritten desktop-app code was merged) through v0.22.0, with no commits in that range touching the vulnerable functions.”

The latest Ollama release (v0.23.0) shipped two days ago, again with no patch, he told us.

Help Net Security has reached out to the Ollama maintainers to ask when they plan to push out fixes, but haven’t heard back from them.

Until a definitive solution is provided, Dmitruk advises users to switch off the Auto-download updates option in Ollama’s settings.

“Turning it off short-circuits the background download check (updater.go:324) before any update response is fetched, so the path traversal write never happens and nothing lands on disk. Users who want to be thorough should also remove any existing Ollama shortcut from %APPDATA%\Microsoft\Windows\Start Menu\Programs\Startup to disable the silent on-login execution route.”

Subscribe to our breaking news e-mail alert to never miss out on the latest breaches, vulnerabilities and cybersecurity threats. Subscribe here!