The Safety Case Physical AI Hasn't Finished Writing

The fence is gone. The proof is not.

Humanoids and mobile robots are moving into spaces people already occupy, which means the old safety premise, keep the robot and the person apart, no longer holds by design. One recent industry analysis of U.S. shipbuilding and defense automation describes humanoids as built to “fit directly into existing workspaces previously used by people,” requiring safety functions that detect workers in real time and autonomously stop or reroute economy.ac. Denmark’s The Mobile Robot Company built its CE-marked vehicle on similar logic, pairing an established forklift manufacturer’s chassis and motor engineering with autonomy layered on top beaconjournal.com.

That a cyber vulnerability can now be a machine-safety hazard is not a new regulatory idea. The EU Machinery Regulation already directs ISO 12100 risk assessments to treat “corruption of safety-related hardware, software, and data as a hazard source,” with IEC 62443 threat modeling supplying the attack scenarios that Sections 1.1.9 and 1.2.1 require designs to withstand jamasoftware.com. What’s unresolved is whether anyone has actually closed that loop for a machine that shares floor space with a person, rather than a machine behind a fence talking to a plant network.

What the evidence actually shows

The available proof points are vendor-supplied, and buyers should treat them that way. Techman Robot’s TM AI Cobot S Series earned IEC 62443-4-2 certification, which is a genuine, independently audited standard, but it covers a collaborative arm, not a mobile or humanoid platform operating without a cage prnewswire.com. TMRC’s CE marking speaks to chassis and motor safety engineering; its press material does not establish that IEC 62443 threat modeling was a first-class input to that process beaconjournal.com. Neither case demonstrates the specific, harder claim: that a robot in physical contact range has been assessed against adversarial interference the way it has been assessed against mechanical failure.

There’s a second gap worth naming. A security vendor’s own analysis argues that assessments are shifting from static audits to “a living picture of where the organization stands” shieldworkz.com. That is a sales pitch for continuous monitoring as much as it is a description of market practice, and it implicitly concedes that most certifications today are still frozen at the audit date. Meanwhile, a separate line of analysis on AI platform reliability warns that safety controls which act autonomously, stopping a process or rerouting a path, become a new kind of dependency: one whose failure mode is opacity, where the system withholds the reason it intervened cybrsecmedia.com. For a humanoid that stops near a line worker, an unexplainable stop is not a minor UX problem. It is the safety function itself behaving as an unverified black box.

What buyers should actually ask for

Point-in-time certification and CE marking answer the mechanical and electrical questions. Neither answers whether the vendor has threat-modeled the specific scenario of a person in contact range with a compromised actuator, and neither proves the assessment stays current after deployment. Safety switch and HMI vendors at recent trade shows are still solving the fenced-robot problem well smartindustry.com. Nobody yet has published evidence that the unfenced problem has been solved the same way.

Ask the vendor to show the threat model, not just name the standard it was tested against

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The central argument—that certifications for fenced robots don’t prove safety for unfenced human-contact scenarios—is coherent and well-supported, though the claim that ‘nobody yet has published evide
Source & Claim Verification	Qwen · hosted fallback (qwen-2.5-72b-instruct)	cleared. All factual claims are traced to citations, but some sources are not directly linked to specific claims, which could be improved for clarity.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately reflects key requirements of ISO 42001, the EU AI Act, and NIST AI RMF regarding risk assessment and continuous monitoring but lacks explicit mapping to 21 CFR Part 11 and EU G
Technical Accuracy	Llama	cleared. The article accurately highlights the gap in safety assessments for unfenced robots and the need for threat modeling and continuous monitoring, but could be improved with more technical depth on speci
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively identifies and critiques vendor-supplied evidence, consistently highlighting the gap between current certifications and the specific safety claims for unfenced physical AI.
Novelty & Non-Duplication	Grok	held. The unfenced-robot / cyber-as-safety-hazard gap is a competent stitch of recent vendor PRs (Techman 62443, TMRC CE, shipbuilding humanoids) and standard EU Machinery/IEC talking points, not a wire dup
Validation	DeepSeek	cleared. The briefing correctly shows no vendor-supplied case demonstrates a complete, adversarial threat model for a mobile robot in physical contact range, but it fails to refute the central claim as no prov

Sources cited: 20. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.