CVE-2026-52394: DeepSeek V4 incorrect access control in chat interface

CVE-2026-52394 · Severity: critical · CVSS 9.1 · Published 2026-10-13

DeepSeek V4 is an AI chat service that assists users with a wide range of tasks. A critical flaw in the chat interface allows an unauthenticated attacker to craft a malformed message that bypasses session isolation, exposing internal reasoning chain data from other users' conversations. This could leak sensitive information, prompts, and proprietary problem-solving steps from concurrent users.

Vendors: DeepSeek.

Quick answers

Is CVE-2026-52394 exploited in the wild?
No. Junglewise Threat Intelligence has no report of CVE-2026-52394 being exploited in the wild.
How severe is CVE-2026-52394?
Critical severity, with a CVSS score of 9.1, published 13 October 2026.
Which products and versions are affected?
DeepSeek V4.
Is there a fix for CVE-2026-52394?
A patch was published: Vendor has fixed the vulnerability Until you apply it: Implement strict session isolation for the reasoning chain buffer at the inference server level, validate and sanitize the thought trigger format before processing, and ensure each inference request is scoped to its own session context. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.
How is CVE-2026-52394 attacked?
It is attacked over the network, in a web application.
Am I affected by CVE-2026-52394?
Check whether you run DeepSeek V4. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

How to fix

Implement strict session isolation for the reasoning chain buffer at the inference server level, validate and sanitize the thought trigger format before processing, and ensure each inference request is scoped to its own session context.

Technical details

DeepSeek V4's chat interface (chat.deepseek.com) fails to properly isolate reasoning session buffers across different users. By sending a crafted input with malformed thought triggers (specifically the sequence "?? <think ????????????"), an unauthenticated attacker can cause the system to return reasoning chain content—including prompts and inference outputs—from other active or cached user sessions. The vulnerability is a multi-tenant session isolation failure (CWE-201, CWE-209) exploitable via network-based direct API calls; the vendor has issued a fix.

Affected products

  • DeepSeek V4

Timeline

  • 2026-07-24: disclosed
  • 2026-10-13: advisory
  • patched: Vendor has fixed the vulnerability

References

This page as Markdown. Free to reuse under CC BY 4.0 with a link back.