Vulnerability  ·  2026-11-06

Jan agent tools workspace-guard bypass: prompt injection writes files outside workspace via dangling symlinks

VulnerabilityMedium impactGlobalCVE-2026-122235
The agent-tools gate.rs guard can be bypassed via dangling symlinks: when the agent follows a symlink planted in the repository, the write/edit tools write outside the sandboxed workspace. A prompt-injection payload steers the agent into creating files such as an authorized_keys entry.
Agent workspace guards are a primary containment control against prompt-injection exfiltration. A symlink bypass converts an injected prompt into host file writes, which can be escalated to persistence (e.g., SSH key) on the developer machine running Jan.
An attacker supplies a malicious repository; prompt injection makes the agent's write/edit tools traverse a dangling symlink so files are written outside the workspace — for example, to ~/.ssh/authorized_keys on the host.
Jan >= 0.8.5, < 0.8.8
Upgrade Jan to >=0.8.8 (GHSA-7j4g-mj5j-p842 / commit 840c33e6).
NVD CVE-2026-122235Jan advisory GHSA-7j4g-mj5j-p842Jan fix commit
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →