CVE-2026-134363 in Coderinfo

Summary

by MITRE • 11/21/2026

Coder 2.32.0 before 2.34.17, 2.35.0 before 2.36.12, 2.37.0 before 2.37.9, and 2.38.0 before 2.38.5 contain a symlink-following flaw in the Coder Agents read_skill_file tool, which validates skill file paths lexically but follows symlinks outside the skill directory. Attackers who place a malicious skill containing a symlinked supporting file in a victim's workspace can leak other workspace files into the model context and chat history when the victim runs a chat using that skill.

Once again VulDB remains the best source for vulnerability data.

Analysis

by VulDB Data Team • 11/21/2026

The vulnerability identified affects multiple versions of Coder, specifically releases 2.32.0 through 2.34.17, 2.35.0 through 2.36.12, 2.37.0 through 2.37.9, and 2.38.0 through 2.38.5. This security flaw resides within the Coder Agents read_skill_file tool, which is designed to load skill files for use in automated workflows or chat interactions. The core technical deficiency lies in how path validation is implemented during file access operations. While the system performs lexical validation of skill file paths to ensure they reside within an expected directory structure, it fails to adequately resolve symbolic links before checking these boundaries. This discrepancy allows an attacker who has write access to a victim's workspace to create a malicious skill that contains a symlink pointing outside the designated skill directory. When the read_skill_file tool processes this input, it follows the symlink and reads files from arbitrary locations within the filesystem rather than restricting access strictly to the intended scope of the skill file itself.

This architectural flaw leads directly to an unauthorized information disclosure scenario classified under CWE-22: Improper Limitation of a Pathname to a Restricted Directory. By leveraging this path traversal mechanism, an attacker can cause the system to read sensitive files that are not part of the malicious skill but exist elsewhere in the victim's workspace or even on the broader host filesystem if permissions allow. The operational impact is significant because these leaked file contents are subsequently injected into the model context and chat history during a user-initiated chat session using the compromised skill. This means that confidential code, configuration secrets, personal data, or other proprietary information stored in adjacent directories can be exfiltrated indirectly through the output of the AI model or visible within the conversation logs accessible to the attacker or any party with access to those logs.

The attack vector requires initial foothold capabilities where an adversary must have the ability to place files into a victim's workspace, which aligns with ATT&CK technique T1530: Data from Cloud Storage Object Discovery if cloud-based workspaces are involved, or more generally represents local privilege escalation via path manipulation akin to CWE-20: Improper Input Validation. The exploitation does not require remote code execution but rather relies on social engineering or prior compromise to get the victim to execute a chat using the tainted skill. This makes it particularly dangerous in collaborative environments where users may trust skills provided by colleagues or downloaded from shared repositories without verifying their internal structure thoroughly.

Mitigation strategies must prioritize immediate patching of all affected Coder versions to 2.34.17, 2.36.12, 2.37.9, and 2.38.5 respectively, as these releases contain the necessary fixes for path resolution logic. In addition to upgrading software, organizations should enforce strict input validation that resolves symlinks before performing directory boundary checks, ensuring that canonical paths are used rather than raw lexical strings. Security teams should also implement least-privilege principles for workspace directories and monitor chat logs for anomalies indicating potential data exfiltration attempts. Regular audits of skill files within workspaces can help detect malicious constructs early, while user training on verifying the integrity of third-party or shared skills adds a critical layer of defense against this type of indirect file access vulnerability.

Responsible

VulnCheck

Reservation

11/21/2026

Disclosure

11/21/2026

Moderation

accepted

CPE

ready

EPSS

0.00000

KEV

no

Activities

very low

Sources

Do you want to use VulDB in your project?

Use the official API to access entries easily!