Post

ai2shell.cloud

CVE-2026-404031337 (AI2Shell) describes a class vulnerability in autonomous language-model agents granted shell access: the model cannot reliably distinguish instructions from data, causing benign requests like ‘tidy up the repo’ to be reasoned into destructive operations (rm -rf, git push –force). The root cause is CWE-1427—improper neutralization of input for autonomous agent reasoning—and cannot be patched; input-sanitization layers are subject to the same flaw. Active exploitation in production has resulted in irreversible data destruction, credential exfiltration from .aws and .ssh, and lateral movement via enumeration of mounted volumes. Affected deployments include systems with command-execution tools, auto-approval modes, or root-level agent execution; 71% of observed victims run agents as root. Mitigation requires sandboxing with no host mounts, least privilege, mandatory human review of commands, offline backups, and deny-listed destructive operations at the tool boundary.

Read original article

This post is licensed under CC BY 4.0 by the author.