CVE-2026-107802: SumatraPDF command-line argument injection in AI translation

CVE-2026-107802 · Severity: info · Published 2026-10-08

SumatraPDF is a Windows document reader that includes AI-powered translation features using external services like Grok, Claude, and OpenAI Codex. When translating selected or pasted text, the application fails to properly escape special characters before passing the text to these AI backend tools, allowing an attacker to inject command-line flags that could bypass safety controls or redirect the request to a malicious AI model. This requires the user to select or paste attacker-controlled text and trigger the translation feature with an installed AI backend.

Technologies: SumatraPDF. Vendors: SumatraPDF.

Quick answers

Is CVE-2026-107802 exploited in the wild?
No. Junglewise Threat Intelligence has no report of CVE-2026-107802 being exploited in the wild.
How severe is CVE-2026-107802?
Informational severity, published 8 October 2026.
Which products and versions are affected?
SumatraPDF 3.6.1 and earlier.
Is there a fix for CVE-2026-107802?
Yes: fixed in No fixed version identified as of the advisory date; patch commit 6b88a751 adds proper argv quoting via QuoteCmdLineArgTemp.. Until then: Do not translate untrusted or attacker-controlled text using the AI translation feature until the application is updated. If an AI backend is not installed, the translation feature cannot be exploited. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.
How is CVE-2026-107802 attacked?
It is attacked with local access, in a software library.
Am I affected by CVE-2026-107802?
Check whether you run SumatraPDF at an affected version. 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

  • Fixed in: No fixed version identified as of the advisory date; patch commit 6b88a751 adds proper argv quoting via QuoteCmdLineArgTemp.

Do not translate untrusted or attacker-controlled text using the AI translation feature until the application is updated. If an AI backend is not installed, the translation feature cannot be exploited.

Technical details

The vulnerability is an argument injection (CWE-88) in src/SelectionTranslate.cpp's BuildGrokTranslateCmdLineTemp(), BuildClaudeTranslateCmdLineTemp(), and BuildCodexTranslateCmdLineTemp() functions. These functions use incomplete escaping (only " → \") when embedding user-controlled text into Windows CreateProcessW command lines, but proper Windows argv quoting requires backslashes before quotes to also be escaped. Text ending in \" breaks out of the quoted argument and allows injection of additional flags like --always-approve or --model. The attack requires local access, an installed AI backend, and user interaction to translate attacker-supplied text, but enables privilege escalation through safety-bypass flags in agentic CLIs.

Affected products

  • SumatraPDF SumatraPDF 3.6.1 and earlier

Timeline

  • 2026-07-28: disclosed
  • 2026-10-08: advisory

References

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