CVE-2026-87986

Mistral Vibe, Mistral AI's agentic coding CLI that executes shell commands on a developer's machine, can be bypassed because commands containing shell constructs its parser cannot interpret are not fully inspected against its permission rules. When the agent builds or runs such a command, the uninterpreted portions are omitted from the command shown for inspection, so embedded commands execute on the user's system without user approval. An attacker who can influence the content the agent acts on (for example task instructions, repository content, or model output) gains arbitrary code execution with the victim user's privileges; CVSS 4.0 rates this 10.0 (Critical). All users of Mistral Vibe are potentially affected, but the available data does not specify affected version ranges. No exploitation has been reported: the flaw is not in CISA's KEV, and no public proof-of-concept is known.

What to do: Update Mistral Vibe to the latest available release and monitor Mistral's advisory/GitHub channel for the patched version, since no fixed version is specified in the available data. Until patched, limit Vibe to trusted repositories and content, run it with least-privilege credentials or inside a sandbox/container, and remember that approval prompts may not show every part of the command being executed. Check endpoint logs for unexpected child processes or shell invocations spawned by Vibe.