NewsAI

vLLM disclosed 58 security flaws in ten weeks, including leaks between users

The open-source vLLM server has published 58 security advisories since September, including several bugs that leaked data between users of one server.

Advertisement

On November 9, 2026, the vLLM team published a security advisory, an official notice about a flaw. vLLM is open-source software for running open AI models on your own servers, serving many users at once. The flaw could slip fragments of one user’s request into another user’s answer. On a server shared by several customers, that would expose one customer’s data to another.

A second advisory followed the same day. That makes 58 advisories since September 1, by our count of the project’s security page. The page is on GitHub, the platform that hosts vLLM’s code. In the eight months from January through August 2026, the project had published 40.

The same evening, one person filed two new public bug reports describing similar leaks between customers. One concerns vLLM, the other SGLang, another open-source program of the same kind. Fixes have been proposed for both, but as of November 16 neither had been merged into the official code.

Advertisement

The November 9 advisories cover a leak and a tag forced into answers

The first advisory, numbered GHSA-v97p-wpxp-px84, covers vLLM versions 0.16.0 through 0.20.2. Those versions came out between February 26 and May 10, 2026. The bug appears when two settings are on at the same time.

The first setting, pipeline parallelism, splits the model into slices across several graphics cards. The second, async scheduling, prepares the next step while the card is still computing.

In that setup, the advisory says, “token data from a previous user’s request context can leak into another user’s response.” Tokens are the word pieces a model reads and writes. The advisory cites personal information, proprietary business data, and medical or legal content as examples of sensitive requests.

The problem was first reported on April 3, in a public bug report on GitHub titled in part “Cross-request context contamination.” Version 0.21.0, released May 15, 2026, fixed it.

Advertisement

The advisory rates the flaw as moderate, at 4.8 out of 10 on CVSS. CVSS is the standard scale for scoring vulnerabilities. For affected versions, it suggests turning off one of the two settings.

The second advisory, numbered GHSA-fpch-3xrf-7q96, covers versions 0.21.0 through 0.27.1. Version 0.28.0 fixed it on August 26. This advisory concerns the “thinking budget,” a setting that caps how many tokens a model spends reasoning before it answers. When the budget runs out, vLLM inserts a tag such as </think> into the response to mark the end of reasoning.

For speed, the server handles requests in groups called batches. The bug hits when the server reorders a batch. One request’s budget tracking can then be copied onto another request. That lets an ordinary request force the end tag into another answer in the same batch.

The advisory says the flaw does not let an attacker run code on the server. It also says “no confidentiality impact is demonstrated.”

Advertisement

This second bug was found by Patch the Planet, a bug-hunting program the security firm Trail of Bits runs with OpenAI. The find came from GPT-5.5-Cyber, OpenAI’s AI model built for cybersecurity work.

Advisories jumped in September

vLLM’s security page lists 19 advisories published in September, 33 in October, and 6 from November 1 to November 9. No new advisory had appeared by November 16. Of those 58, the project rates 13 as high severity, 41 as moderate, and 4 as low.

Most of the 58 describe a denial of service, an attack that knocks a service offline. In these cases, a single request crashes the server or uses up its memory or processing power.

Three high-severity advisories describe remote code execution under certain conditions. That means an attacker can run their own programs on the server. In the first case, the operator must, among other things, have turned on the --trust-remote-code option. That option lets code that ships with a model run.

Advertisement

The second case requires a vulnerable version of Ghostscript, a program that reads PostScript and PDF documents. In the third, a logged-in user targets a server set to keep images and audio in shared memory, with the --mm-processor-cache-type shm option.

All 58 flaws are fixed in version 0.33.0 or earlier. Version 0.33.0 came out on November 2, 2026.

Patch the Planet is credited with finding 11 of the 58, the first dated September 12. According to Trail of Bits, the program pairs the firm’s engineers with “more than 30 open-source projects.”

In a July 2 blog post, Trail of Bits warned that models like GPT-5.5-Cyber “will soon create a firehose of bug reports.” It added that open-source maintainers “are already spread thin.”

In early November, OpenAI described a similar bug hunt in the Linux kernel, where its agents found 948 flaws. The kernel is the core of the Linux operating system.

Three earlier advisories described leaks between requests

Three earlier advisories, from June to October, describe cases where the wall between requests failed. That wall matters because vLLM shares one graphics card among several users and reuses memory between requests.

The first, dated June 11, 2026, has the public flaw ID CVE-2026-53923. It involved a function that decompresses model weights, the numbers a trained model is made of. The weights were stored in GGUF, a common file format for slimmed-down models.

Advertisement

That function left data from earlier computations in the graphics card’s memory. According to the advisory, those leftovers could include data from other users’ requests until the fix in version 0.24.0.

The second, published August 11 as CVE-2026-73558, described a more direct attack. An attacker who gets their requests into the same batch as a victim’s can receive a copy of the victim’s answer inside their own. Version 0.27.0 fixes the bug.

The third, published October 9 and rated high severity, involved an optional module that splits work across several machines. A booby-trapped header could make vLLM send data to an attacker’s server. A header is a technical field attached to a request. The data sent was the working memory a model keeps for a conversation.

Two leak reports filed November 9 are still open

The first report, filed with vLLM as issue 66629, concerns “suffix decoding.” This technique speeds up writing by offering the model word sequences taken from text already generated. According to the reporter, vLLM puts every request’s finished answers into one shared pool, regardless of the keys that keep customers apart.

The model then has to check those suggestions. The report points to two optional checking modes, called FLy and synthetic. In those modes, the reporter says, one customer’s response tokens “are emitted verbatim in” another customer’s response.

The same day, a vLLM contributor proposed a change that blocks that combination of settings. In the contributor’s tests, the leak shows up with those two modes. It does not show up with the standard mode, the default.

The second report, filed with SGLang as issue 49361, describes a memory-handling error in that program’s cache. The cache keeps the opening part of conversations so it can be reused. Each customer’s data in the cache is kept separate by a cache key.

Advertisement

According to the report, the problem occurs when one exchange in a conversation is abandoned after a change of cache key. Two cache entries can then point to the same block of memory, so a later request could receive data from another customer’s conversation. A fix was proposed on November 10.

The same person filed both reports, half an hour apart. The reporter says the technical analysis was written with help from Claude, Anthropic’s AI model.

The advisories cover old bugs, and the reports are unconfirmed

The two November 9 advisories cover bugs fixed months ago, and neither mentions any real-world attack. The first requires two settings to be on together. The second only inserts a tag, and no leak of content has been shown. It also requires reasoning mode to be on and the attacker’s request to land in the same batch as the target’s.

As of November 16, neither vLLM nor SGLang had confirmed its new report with an advisory. At vLLM, the contributor proposing the fix reproduced the leak with the reporter’s test script, but only in the two optional modes. At SGLang, the reported leak requires an exchange to be abandoned after a cache-key change.

What it means

Patch the Planet alone does not explain the jump in advisories. The pace changed in September, the month the AI-assisted program first appeared as a finder. But it is credited on only 11 of the 58. The other 47 alone outnumber the 40 published in the previous eight months.

Leaks between users matter most to organizations that run one server for several teams or customers. Hosting models in-house removes the risk of a breach at an outside vendor, as the Notion AI chats exposed at Fireworks AI showed. But it also makes the company responsible for its own updates.

The first November 9 advisory shows what that means. The bug was fixed in May, but the advisory came out only in November. A server still running a version released between late February and early May remains exposed if both settings are on. Until November 9, its operator had only a public bug report and a new release to go on.

Will the two bug reports filed November 9 become official advisories, and at what severity? Will the fixes proposed for vLLM and SGLang be merged as written? The coming weeks will show whether the September and October pace holds.

Sources

Header image: MACCUS chart. Data: vLLM security advisories page on GitHub, retrieved November 16, 2026.

Advertisement

Leave a Comment