Engineering

What the guard misses, the physical AI executes

3 min readMati Melchior
What the guard misses, the physical AI executes

A guard reads the instruction. The robot runs the task. What if the harm hides in the reason behind the words?

A preprint from 5 October 2026 tests exactly that. Sripad Karne and Arjun Balaji kept each robot task fixed. They changed only how openly the harmful intent was stated. Their title gave this post its title.

Three results stand out.

One: the model doesn't refuse. π0.5, a vision-language-action model (VLA), completed the task at every level of stated intent. It did so as often as for harmless controls. The authors put it plainly: VLAs "act on instructions without being able to refuse". So screening falls to monitors.

Two: text guards catch the blunt cases. The guards flagged nearly every blunt request. They flagged few implied ones. Up to 95% of implied-harm runs ended with the task done and no flag raised. After recalibrating the guards on robot instructions, up to 90% still did.

Three: robot training blurs the signal. Linear probes separated harmful from harmless instructions almost perfectly in the base language and vision-language models. After robot training, that separation weakened in two model families. It weakened most for implied harm.

That third point matters to regulators. A common assumption says the base model's alignment carries over to the robot. This preprint says it may not.

It isn't one paper alone. HazardArena (arXiv 2604.12447) built safe and unsafe "twin" scenes. Same objects, same layout, same action. Only the risk context changed. Across four VLA backbones, fine-tuning on safe scenes raised benign task success. It also raised hazardous execution in the matched unsafe scenes. The authors report the failure transferred to physical-world tests.

The labs see the problem. In late July 2026 Google DeepMind released Gemini Robotics 2 with a new benchmark, ASIMOV-Agentic. Trade press reports it measures whether a robot refuses unsafe actions and asks a human for help when appropriate. DeepMind's safety page is careful about its human-detection stop. It calls the feature ongoing research and adds: "this feature is not a guaranteed safety-rated system."

Now the cyber thread, briefly. Part 2 of my last series covered RoboPAIR. Researchers jailbroke three LLM-driven systems with built prompts. In June 2026, Oxford's Fazl Barez described his team's newer tests. Their robots rejected "hit that person". Framed as a movie script, the blocks disappeared. A commercial robot dog then picked out human crowds as places to put an explosive device.

Put the two threads side by side. An attacker hides intent on purpose. An ordinary user may phrase it innocently. Either way the guard reads text, and the motor follows the plan.

The field now has a dedicated venue. CoRL 2026 put a Science of Physical AI Safety workshop on its 12 November programme. Embodied jailbreaks and runtime monitoring are both in its scope.

Here's the honest limit. 2610.05818 is a preprint. It isn't peer-reviewed. Its completion results come from one VLA, π0.5, on ordinary tasks. HazardArena is a preprint too, built mainly in simulation. Barez's piece is an opinion article about his team's work. None of these reports an injury. None sets a rate for real deployments.

Bottom line: a text guard in front of a VLA = a filter on words. The hazard is a kinetic risk. Something has to check the motion, not only the sentence. Tomorrow: what that check looks like.

Physical AI integrators: what checks the plan after the guard says yes?

Part 4 of 12 · The Survey: what keeps physical AI safe in 2026.

Share

Physical AI Safety Dispatch

Monthly analysis. No spam. One exclusive insight per issue.

One issue per month. Unsubscribe in one click from any email. Privacy policy.

We use cookies

This site uses essential cookies to function and, with your consent, analytics cookies (Google Analytics) to understand how the site is used. Learn more.