Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-107802 Overview
CVE-2026-107802 affects SumatraPDF, a multi-format document reader for Windows. Versions 3.6.1 and earlier contain an argument injection flaw in src/SelectionTranslate.cpp. The code embeds selected or pasted translation text into quoted Windows command lines using incomplete quote-only escaping. Attacker-controlled text passed to BuildGrokTranslateCmdLineTemp(), BuildClaudeTranslateCmdLineTemp(), or BuildCodexTranslateCmdLineTemp() can inject additional CLI flags. These flags may alter the model, working directory, approval behavior, or sandbox settings of agentic AI command-line backends when installed. The weakness is classified as [CWE-88] Improper Neutralization of Argument Delimiters.
Critical Impact
Local attackers can inject arbitrary command-line arguments into agentic AI CLIs invoked by SumatraPDF, bypassing sandbox and approval controls of the underlying backend.
Affected Products
- SumatraPDF 3.6.1
- SumatraPDF versions prior to 3.6.1
- SumatraPDF integrations with Grok, Claude, and Codex agentic CLI backends
Discovery Timeline
- 2026-10-08 - CVE-2026-107802 published to NVD
- 2026-10-09 - Last updated in NVD database
Technical Details for CVE-2026-107802
Vulnerability Analysis
The vulnerability resides in SumatraPDF's AI translation feature. When a user selects text or pastes content for translation, the application constructs a Windows command line to invoke an external agentic CLI backend. The builder functions apply only quote escaping, not full argv-safe quoting, before embedding user-controlled text. Windows command-line parsing treats unescaped spaces, backslashes, and quote sequences as argument separators. An attacker who controls the translated text can therefore terminate the intended argument early and append flags of their own choosing. Because the invoked backend is an agentic AI runner, injected flags can redirect the working directory, select a different model, auto-approve actions, or disable the sandbox.
Root Cause
The root cause is incomplete command-line escaping in BuildGrokTranslateCmdLineTemp(), BuildClaudeTranslateCmdLineTemp(), and BuildCodexTranslateCmdLineTemp(). These functions wrap input in quotes but do not apply the Microsoft CommandLineToArgvW quoting rules required for CreateProcessW. The upstream patch introduces a dedicated QuoteCmdLineArgTemp helper and reclassifies the input field as raw user data that providers must escape before use.
Attack Vector
Exploitation requires local access, user interaction, and a vulnerable agentic CLI backend installed on the system. An attacker crafts a document or clipboard payload containing escape sequences. When the user selects the text and triggers translation, the malicious tokens are concatenated into the command line and parsed as flags such as --cwd, --model, --yolo, or sandbox-bypass switches supported by the backend.
Str sessionId;
Str filePath; // document the user is reading
Str dir; // cwd for the process
+ // Raw user input (not pre-escaped). Providers must pass it through
+ // QuoteCmdLineArgTemp when embedding it in a CreateProcessW command line.
Str escapedInput;
int option = 0; // effort / sandbox index
bool flag = false; // skip permissions / always approve / skip sandbox
Source: GitHub Commit 6b88a751
Stop Remote Code Execution Before It Runs
Singularity™ Endpoint detects and blocks exploit attempts at machine speed, including zero-days and novel attack techniques.
Stop Exploit AttemptsDetection Methods for CVE-2026-107802
Indicators of Compromise
- Child processes of SumatraPDF.exe launching grok, claude, or codex CLI binaries with unexpected flags such as --cwd, --model, --yolo, --dangerously-skip-approvals, or --no-sandbox.
- Agentic CLI sessions whose working directory differs from the SumatraPDF process directory.
- Clipboard or document content containing stray double-quote and backslash sequences preceding flag-like tokens.
Detection Strategies
- Monitor process creation events (Windows Event ID 4688 or Sysmon Event ID 1) for SumatraPDF.exe spawning known agentic CLI executables and inspect the full command line.
- Alert on command-line patterns where translation payload text is followed by -- or - style flags.
- Baseline normal SumatraPDF-to-CLI invocations and flag deviations in model, directory, or approval arguments.
Monitoring Recommendations
- Ingest endpoint process telemetry into a SIEM and correlate SumatraPDF parent processes with agentic CLI child processes.
- Enable Windows command-line auditing via Group Policy to capture full argv strings.
- Track installation of Grok, Claude, and Codex CLI tooling to scope exposed hosts.
How to Mitigate CVE-2026-107802
Immediate Actions Required
- Disable or uninstall the Grok, Claude, and Codex CLI backends on hosts running SumatraPDF 3.6.1 or earlier until a fixed release is available.
- Instruct users to avoid using the AI translation feature on untrusted document content or clipboard data.
- Restrict SumatraPDF usage to standard user accounts without privileged workspace access.
Patch Information
No fixed SumatraPDF release is available as of this review. An upstream code fix is present in commit 6b88a751ab34a28bafca3bdd635aaa2f4c1e06b6, which introduces proper argv quoting through a new QuoteCmdLineArgTemp helper and routes provider inputs through it. Refer to the GitHub Security Advisory GHSA-xvxg-cwmx-hr7j for the vendor's advisory and the upstream commit for the code change.
Workarounds
- Remove or rename the agentic CLI executables (grok, claude, codex) so SumatraPDF cannot launch them.
- Configure application control policies to block SumatraPDF from spawning AI CLI binaries.
- Disable the AI translation feature in SumatraPDF configuration where possible.
# Example WDAC/AppLocker-style rule concept: block SumatraPDF from launching AI CLIs
# Parent: SumatraPDF.exe
# Child: grok.exe, claude.exe, codex.exe
# Action: Deny
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

