
Key Takeaways
- Physical AI security has no clear owner in the enterprise. It falls between AI safety, traditional cybersecurity, and robotics engineering, so no single playbook covers a system that perceives, decides, and acts.
- Agents now produce physical actions. Researchers manipulated three LLM-controlled systems, often with a 100 percent success rate, and harm in the physical world depends on context that prompt-level controls can’t see.
- The Model Hardware Standard (MHS) is one signal of a wider shift. Agentic operational technology (OT) and agentic Internet of Things (IoT) are removing the integration friction that quietly protects equipment.
- The Purdue Model governs where agents connect, not what they do. Segmentation stays essential, but an agent’s logical reach can span levels even when its network access doesn’t, so authorization has to move to the action.
- Hard boundaries are the common control. Deterministic limits below the agent, plus action-level authorization at the execution gateway, hold regardless of model behavior.
Physical AI security has an ownership problem. The AI safety community studies model behavior, cybersecurity teams protect data, networks, and endpoints, and robotics engineers build for reliability, yet a system that perceives, decides, and acts in the real world belongs fully to none of them. That gap has always been there, but AI amplifies the risk for every security leader who runs plants, labs, fleets, or connected devices.
The fundamental components are arriving concurrently. Humanoid robots are moving beyond pilots, agentic AI is connecting to operational technology (OT) and Internet of Things (IoT) estates, and in August, Anthropic previewed the Model Hardware Standard (MHS), a specification that lets agents operate physical equipment through a common interface. MHS is one signal among several, but it makes the stakes tangible.
Physical AI Security Sits Between Three Disciplines
The World Economic Forum put the problem plainly in July 2026: security for agentic AI in physical devices falls between AI safety research, traditional cybersecurity, and robotics engineering. The same piece notes that the National Institute of Standards and Technology (NIST) AI Agent Standards Initiative, which is defining execution isolation, least privilege, and auditability for software agents, doesn’t yet cover robots, autonomous vehicles, or any other physical systems.
The evidence that this gap is real isn’t hard to find. According to the same report, a team at a Capgemini lab spent a week inside a newly purchased humanoid robot and found an onboard computer they couldn’t inspect, persistent connections to servers in the robot’s country of origin, and a Bluetooth flaw that was already publicly cataloged. Separately, researchers examining Unitree humanoids rooted the robots and found a way to reach any unit within Bluetooth range.
Two properties set physical AI apart from the systems most teams already defend. Control loops run in milliseconds, so a delay that costs an IT application a moment can cost a robot its balance. And consequences leave the network: a compromised system can move, heat, cut, or open something.
When Prompts Become Physical Actions
Agentic AI security for software has already taught us that soft guardrails are probabilistic. Researchers at the University of Pennsylvania showed what that means once a model has a body. Their RoboPAIR attack bypassed the safety controls of three LLM-controlled systems (a self-driving model, a wheeled robot with a GPT-4o planner, and a robot dog), often with a 100 percent success rate. IEEE Spectrum reported that a prompt casting the request as a scriptwriting exercise was enough to bypass safety filters.
The finding I’d put in front of every security architect is about context. As the researchers explain, whether a robotic action is harmful depends on its surroundings: a command to walk forward is harmful if a person is in the path and benign if not. A prompt filter can’t evaluate a fact about the physical world it can’t see. Detection after exfiltration is not sufficient, and detection after a collision is catastrophic.
MHS Signals the Path Forward
MHS is the clearest example of where the interface layer is heading. It’s a shared specification for letting AI agents operate physical devices safely, developed with the Howard Hughes Medical Institute’s Janelia Research Campus, and AI agents reach it through the Model Context Protocol (MCP), a command-line interface, or code. MHS targets labs and advanced manufacturing today, but analysts are already reading it as an IoT control plane. That’s the right lens. Whether the device is an automated liquid handler, a programmable logic controller (PLC), a sensor array, or a robot arm, the pattern is the same: an agent reasons about state, plans, and acts through a standard interface.
Until now, connecting and orchestrating physical equipment has meant proprietary protocols, fragile custom scripts, and one-off integrations. That friction kept the work expensive and limited it to well-funded projects, and it forced a review every time. MHS lowers that complexity, so connections that were never worth building become routine, and they can be built in software with little oversight. Integration will scale faster than governance, which is why security has to be designed before the pilot, not after the incident.
Agents also work in ways our controls weren’t built to watch. In the MHS work, Claude adjusted a laser, watched the result on a camera, and iterated before packaging what it learned into a deterministic script. Agent-written scripts become changes to your process, unattended overnight runs remove the human observer, and parallel orchestration lets one bad decision reach several systems. A shared interface also concentrates risk: if many devices are reachable through one agent platform, compromising the agent, its credentials, or one trusted integration can open a route to all of them. Updating a model, driver, or prompt can change behavior even when the machine hasn’t.
Governments are converging on the same message. The Cybersecurity and Infrastructure Security Agency (CISA) and its partners published Principles for the Secure Integration of AI in OT, which covers AI agents explicitly. In May, six agencies followed with Careful Adoption of Agentic AI Services, which tells organizations to assume agents may behave unexpectedly and to favor resilience, reversibility, and containment over efficiency.
Purdue Governs Where Agents Connect, Not What They Do
The Purdue Model will remain relevant, but AI agents challenge one of its fundamental assumptions: that you can secure a system primarily by controlling network communication between layers. In a traditional plant, applications, control systems, and physical processes have well-defined responsibilities. An agent is a new kind of actor that can reason across layers, gather context from several systems, and invoke tools that change the physical world. The important shift is that an AI agent’s logical reach can span the Purdue Model even when its network access shouldn’t.
NIST’s guidance on OT security emphasizes segmentation, isolation, and controlled communications between zones, and agent security has to build on those controls rather than replace them. Purdue tells you how to structure and segment the environment. It doesn’t tell you whether an agent should be trusted to perform a particular action in a given context. The questions shift from which device can talk to which device toward which agent can invoke which capability against which asset, and whether an authorized tool can issue an unsafe but technically valid command. An agent with legitimate access can be manipulated into acting in ways no one intended, and every connection involved would be permitted.
Keeping Purdue for segmentation and adding an agent security control plane across it is a solid approach built on four controls.
- Agent identity and provenance. Identify the executing agent and the authority under which it operates.
- Action-level authorization. Authorize individual tools and operations, not just network sessions.
- Context-aware constraints. Evaluate the target asset, operational state, action limits, and approval requirements.
- Continuous oversight. Monitor actions and delegation, and revoke access without disrupting essential control functions.
The control plane is a logical capability, not necessarily a single product, and enforcement has to hold at the execution boundary even during a loss of connectivity. It complements safety instrumented systems, PLC protections, interlocks, and segmentation; it doesn’t replace them. Three models end up working together: Purdue and the International Society of Automation’s ISA-95 standard for where systems sit, International Electrotechnical Commission (IEC) 62443 zones and conduits for which communications are permitted, and agent security for who or what is acting and whether the proposed action is allowed.
Hard Boundaries Are the Common Control
Across AI systems, robots, instruments, and connected devices, one architectural theme keeps surfacing. Hard boundaries are deterministic, code-level controls where the model never gets a vote. Zenity Labs research makes the distinction plainly: hard boundaries fail through software bugs, while soft boundaries fail by design.
As the stakes increase, enforcement first is not optional. MHS points the same way. The MHS preview says safety limits live in the driver, not the prompt, which means an agent can’t finesse its way past them. The same instinct applies elsewhere. Emergency stops and interlocks belong in hardware and controller logic. Permissions should reflect the device, the command, the operating state, and the potential physical impact, not just whether an agent has network access. The Five Eyes intelligence agencies’ guidance leaves the choice of which actions need human sign-off to system designers, not the agent.
What security leaders can do now
You don’t need a new security program for physical AI. You need to extend the one you have, and the work starts before the first pilot touches production.
- Inventory every path from an agent to a physical system, including robots, instruments, and IoT devices, and classify each as observe, recommend, or execute.
- Keep Purdue segmentation, and treat each agent-to-device path as a conduit that is allow-listed explicitly instead of extending a flat network.
- Keep safety functions independent of the model, with hard limits in drivers, controller logic, and hardware interlocks, and test emergency-stop behavior and the path back to manual operation on a schedule.
- Govern the model, driver, prompt, and device description as one versioned configuration, and review agent-written scripts before they touch production equipment.
Standards and robots are shipping faster than the security practices around them, and the NIST gap noted earlier means nobody is about to hand you the playbook. The agent is the new endpoint, and now it has hands.
All ArticlesRelated blog posts

Vibe-Coded Agents Inherit Their Templates’ Blind Spots
Secure AI agent development is the area where the homegrown, code-first distinction matters most. When an organization...

Every AI Agent Needs Its Own Identity, Not a Borrowed One
AI agent identity is where good intentions most often meet shortcuts. A developer needs an agent to reach a database,...

What Market Conduct Examiners Are Starting to Ask About Claims and Underwriting AI Agents
Agentic AI for life insurance has moved out of the innovation lab and into the decisions regulators already watch...
Secure Your Agents
We’d love to chat with you about how your team can secure and govern AI Agents everywhere.
Get a Demo