al-ice.ai

Claroty Team82 — LLMs Discover Critical ICS Vulnerabilities Without Prior Disclosures

Security by al-ice.ai Editorial

AI relevance: Claroty Team82 demonstrates that a frontier LLM can autonomously rediscover five critical ICS vulnerabilities (three OS command injection, one out-of-bounds write, one XSS — all CVSS 9.8) in a Zenitel video intercom platform, without access to the original vulnerability details — compressing manual research that took hours into a hands-free automated workflow relevant to AI-assisted ICS/OT security assessments.

Claroty's Team82 research team published "Hands Free," documenting their use of Claude Opus 4.6 via Claude Code to independently rediscover vulnerabilities in the Zenitel TCIV-3+ IP video intercom — a device deployed in high-security areas and industrial environments:

  • Original manual research (November 2025): Team82 extracted the firmware, unpacked the UPX-packed ipstweb binary, performed static analysis, and identified five vulnerabilities: CVE-2025-64126 through CVE-2025-64130, covering three OS command injections, one out-of-bounds write, and one XSS — all scored CVSS 9.8.
  • LLM-driven approach: Team82 set up a research working directory with a CLAUDE.md context file and .mcp.json configuration, placing the firmware update in the target folder. The LLM performed the analysis without prior access to the disclosed vulnerability details — discovering the same concrete vulnerabilities independently.
  • The LLM found vulnerabilities that had never been publicly detailed. The original disclosures contained no technical specifics; the model's analysis produced concrete evidence of the flaws on its own.
  • Hands-free vulnerability research is arriving. Team82's experiment mirrors the trajectory of Anthropic's Project Glasswing (Mythos), which has already been demonstrated to shrink vulnerability discovery timelines from months to days at scale.
  • ICS/OT security implications: The target platform manages physical access control to secure areas, with SIP/VoIP and remote management interfaces. Automated AI-driven analysis of ICS firmware lowers the barrier to discovering exploitable flaws in critical infrastructure.
  • Context engineering matters: The CLAUDE.md file acted as the core driver, framing the task as a CTF-style security research scenario and providing extensive guidance on binary analysis approaches.

Why it matters

As frontier models gain the ability to autonomously perform static analysis, unpack obfuscated binaries, and identify command injection and memory corruption flaws in ICS/OT firmware, the window between vulnerability discovery and exploitation will continue to shrink. Defenders can use the same techniques for proactive assessment — but so can attackers with access to the same models. The gap between offensive and defensive AI-assisted research is narrowing.

What to do

  • ICS/OT vendors should treat automated LLM analysis as a baseline assessment — if Claude Opus 4.6 can find your vulnerabilities, attackers with similar access can too.
  • Deploy AI-assisted vulnerability scanning as part of continuous firmware assessment pipelines, not just point-in-time reviews.
  • ICS operators should prioritize network segmentation and access controls for devices like video intercoms that bridge physical security and IP networks.
  • Monitor for the emergence of off-the-shelf "AI vulnerability research as a service" targeting ICS/OT and embedded firmware.

Sources: