All vulnerabilities

Servers and applications

GHSA-3p4q-qm52-c8cp

DoS: one uploaded document blocks the event loop for minutes (unbounded synchronous token counting)

Severity
Medium
Published
17 Nov 2026
Last changed
17 Nov 2026
Added to UpHeron
17 Nov 2026
Fixed in
0.8.8
Weakness
Uncontrolled resource consumption (CWE-400)Regular expression denial of service (ReDoS) (CWE-1333)
Sources
Project advisories (GitHub)

Description

Token counting runs synchronously on the main thread with no bound on input size and no bound on time. Because ai-tokenizer is quadratic in the length of an unbroken character run, a single uploaded document can pin one CPU core and make the entire instance unresponsive until it is restarted.

Any authenticated user can trigger it with no special permissions by attaching an ordinary document. It affects every other user on the instance, not just the caller.

On our shared internal deployment (~750 users) this took production down four times in one day:

  • One CPU core at 100%, HTTP requests unserved, /health timing out at 10s.
  • No recovery without a container restart — the work is not interruptible and there is no timeout.
  • Every in-flight stream for every other user dies with it.
  • Log throughput collapsed ~80x during each incident (~400 lines/min to 5), which is what makes this hard to diagnose from logs: the process looks idle, not busy.
From LibreChat-AI/LibreChat's security advisory. Shortened; the full text is there.

Workarounds

  • Collapse runs of >N identical characters before counting. Measured: 50K underscores goes 2,301ms -> 1ms. The naive version distorts the count (782 tokens -> 3); counting a bounded sample of the run and scaling it linearly stays accurate, because repeated-char runs tokenize linearly (25K -> 392 tokens, 50K -> 782).
  • Cap total input length, above which use the existing /4 estimate. Simplest, but blunt — a 150K-char document with a 100K-char run still costs seconds and would pass a 200K cap.
  • Move tokenization off the main thread.
Workarounds as given by LibreChat-AI/LibreChat's security advisory.

Affected software

SoftwareAffected versions
LibreChatServers and applications · github:librechat-ai/librechat
  • Every version before 0.8.8

About LibreChat

Enhanced ChatGPT Clone: Features Agents, MCP, Skills, DeepSeek, Anthropic, AWS, OpenAI, Responses API, Azure, Groq, o1, GPT-5, Mistral, OpenRouter, Vertex AI, Gemini, Artifacts, AI model switching, message search, Code Interpreter, langchain, DALL-E-3, OpenAPI Actions, Functions, Secure Multi-User Auth, Presets, open-source for self-hosting. Active

From its GitHub repository.
Docker Hub pulls
669,094Docker Hub, 15 Nov 2026
GitHub stars
47,138GitHub, 15 Nov 2026
Forks
9,706GitHub, 15 Nov 2026
License
MITGitHub
First published on GitHub
12 Feb 2023GitHub

References