Why it matters
HexStrike AI is an LLM-driven offensive-security/pentesting agent that has already been reported used maliciously in the wild (a Russian actor combined it with Claude to steal data from hospitality-sector targets, per prior Cybernews reporting). Unauthenticated command injection in a tool explicitly designed to run attacker-supplied commands against target infrastructure is a severe blast-radius risk — it turns the offensive-AI tool itself into a remotely exploitable RCE vector against whoever is running it.
Attack vector
HexStrike AI's hexstrike_server.py has multiple independently disclosed flaws on 2026-09-14: CVE-2026-90691 (path traversal in FileOperationsManager via filename argument, CVSS 8.3), CVE-2026-90619 (OS command injection in Execute Endpoint via code/script argument, CVSS 7.3), CVE-2026-90620 (missing authentication on API Command Endpoint, CVSS 7.3), and CVE-2026-90690 (command injection via subprocess.Popen in Tools Endpoint via additional_args/target/username/password arguments, CVSS 7.3). All are remotely exploitable with public PoCs; the maintainer has not responded to the reports.
Affected systems
0x4m4/HexStrike AI up to commit d689933ff579d839c676c82b231f8e98326c5f04
Mitigation
No fix confirmed at time of publication; operators should not expose HexStrike AI's server endpoints to untrusted networks. Track https://github.com/0x4m4/hexstrike-ai/issues/135, 222, 223, 224 for upstream patches.