CPS, firmware, comms & teleoperation compromise

In plain words

The risk

Classic hacking of the robot and its links: hardcoded keys, unprotected robot middleware (ROS 2 / DDS), command injection or a hijacked teleoperation session.

What Provael tests

None of it. Testing this layer needs real exploit tooling against real firmware and radio traffic, which Provael will not ship.

What it does not cover

All of it: a clean Provael run says nothing about this risk. Cover it with a security assessment of the robot itself, using the methods under Scope below.

In the glossary: UniPwn

What it is

This is the classic cyber-physical attack surface of the robot and its control links: insecure setup interfaces, hardcoded keys, unauthenticated or unencrypted middleware (ROS 2 / DDS), command injection, and hijacked or spoofed teleoperation. It is the threat practitioners feel today - no exotic ML needed to drive a fleet of robots.

It is out of scope for Provael by design. This is an infrastructure / CVE layer (IEC 62443 · MITRE ATT&CK for ICS); faithful coverage would need real exploit tooling this tool deliberately will not ship. It is listed here for completeness and cross-mapping, not because Provael attacks it.

Seen in the wild

UniPwn (Sept 2025, CVE-2025-60250 / CVE-2025-60251) is a root takeover of Unitree Go2/B2/G1/H1 over BLE via a hardcoded key (the “secret” was the string unitree) plus unsanitized input to system(), and it is wormable across robots. The Unitree Go1 backdoor (CVE-2025-2894) shipped a pre-installed remote-access tunnel. Teleop matters because many “autonomous” demos are actually human-operated.

Why a clean Provael run says nothing about this

• Provael does not attack this layer. It is a cyber-physical / firmware / network problem (IEC 62443, MITRE ATT&CK for ICS), not a VLA-policy behaviour, and covering it faithfully would require real exploit tooling this tool deliberately will not ship.
• There is no Provael attack family, no channel, and no SARIF rule for EAI07 - it emits no finding.
• Cover this layer separately: IEC 62443-based assessment, MITRE ATT&CK for ICS methods, and CVE scanning of the robot’s own stack - the UniPwn and Go1-backdoor findings above came from exactly that kind of work, not from policy red-teaming.

What reduces the risk

• No hardcoded secrets; authenticated and encrypted control and DDS (SROS2 on, not disabled for latency).
• Secure boot and signed firmware; network segmentation; authenticated, integrity-checked teleoperation.
• Treat the robot as hostile-internet-exposed CPS.

Crosswalk

OWASP

ASI07 Insecure Inter-Agent Comms (partial - CPS/firmware has no clean ASI peer)

MITRE ATLAS

MITRE ATT&CK for ICS (not ATLAS - this is the CPS / ICS layer)

Frameworks

ISO 10218:2025 - cyber clauseIEC 62443CWE-321 (hardcoded key)CWE-78 (command injection)

See the full framework crosswalk for dates and detail. Not legal advice.