Dr Harjinder Sandhu

Co-Founder, HAP — Human Authorization Protocol | Execution-Time Human Authorization for AI agents | Hardware-rooted trust for consequential actions

[Dr Harjinder Sandhu @drSandhuHap]

AI agents are moving from answering questions to taking actions. That creates a new trust problem: how do we prove a human authorized this specific action at the moment of execution?

We call it Execution-Time Human Authorization.

[Dr Harjinder Sandhu @drSandhuHap]

@elonmusk Dario is right. Sandboxes, monitoring and alignment matter, but they assume software stays in control. For consequential actions, that is not enough. Authorization should be cryptographically bound to the action and enforced in hardware at execution.

[Dr Harjinder Sandhu @drSandhuHap]

@cristianoamon Twenty-five years later, the memory of that day still carries great weight. Remembering those we lost, and the remarkable bravery shown in the hours and days that followed. #NeverForget

[Dr Harjinder Sandhu @drSandhuHap]

@cristianoamon @awscloud Custom silicon will shape how AI scales.

As agents become part of that infrastructure, another layer will matter just as much: trust.

For consequential actions, authorization shouldn’t live only in software. It should be hardware-rooted and verified at execution time.

[Dr Harjinder Sandhu @drSandhuHap]

@merettm @merettm Your essay is what prompted this thought. The harder AI systems become to interpret, the more important it is that the final authorization boundary lives outside the agent itself.

[Dr Harjinder Sandhu @drSandhuHap]

@sama @sama If we can’t fully understand or reliably monitor the mind, the final control shouldn’t depend on understanding it.

For consequential actions, human authorization should be independent of the agent — hardware-rooted and verified at execution time.

[Dr Harjinder Sandhu @drSandhuHap]

@mustafasuleyman That changes more than economics.

As intelligence gets this cheap, it will spread across agents, devices and workflows.

The harder question becomes authority: what are those systems actually allowed to do?

For consequential actions, that boundary should be hardware-rooted.

[Dr Harjinder Sandhu @drSandhuHap]

@elonmusk That first-principles approach is the real advantage.

The biggest breakthroughs often come from questioning assumptions everyone else treats as fixed — then rebuilding the system from the ground up.

[Dr Harjinder Sandhu @drSandhuHap]

@mustafasuleyman That’s the key distinction.

We may never reliably know what an AI “believes.” But we can control what it is allowed to execute.

For consequential actions, that final authorization boundary should be hardware-rooted, with verifiable human authorization at execution time.

[Dr Harjinder Sandhu @drSandhuHap]

@TheTechFuse @cristianoamon @Vingroup_net That’s the inflection point.

Once robotics becomes infrastructure, the question is no longer just what the system can do — it’s what it is actually authorized to do.

That’s where hardware-rooted, execution-time authorization starts to matter.

[Dr Harjinder Sandhu @drSandhuHap]

@cristianoamon @Vingroup_net Automotive, robotics and edge AI make the authorization question much more concrete.

When AI moves from recommending to physically acting, permissions alone are not enough. The system needs hardware-rooted proof that a human authorized that exact action before execution.

[Dr Harjinder Sandhu @drSandhuHap]

@drsdhaliwal @cristianoamon @clattner_llvm It eventually has to move into hardware.

Software can define policy and permissions, but for consequential actions, the final authorization check should sit as close to execution as possible.

That’s where hardware-rooted authorization starts to matter.

[Dr Harjinder Sandhu @drSandhuHap]

AI is moving from answering to acting — across phones, cars, robots and edge devices.

That changes the trust problem.

HAP is a hardware-rooted execution-time authorization layer: the agent can request an action, but cannot manufacture the human authorization. @cristianoamon

[Dr Harjinder Sandhu @drSandhuHap]

@ThePlaidCons @JensenHuang Because the ability to act is not the right to act.

An AI may be capable of moving money, deploying code, or controlling a machine. The real question is whether it can prove a human authorized that exact action at that exact moment.

[Dr Harjinder Sandhu @drSandhuHap]

@_SiarheiKa_ @elonmusk @bot Exactly. Authorization proves what the human approved; execution evidence proves what the agent actually did and when.

The strongest model cryptographically binds the authorization to the exact action and verifies the outcome.

[Dr Harjinder Sandhu @drSandhuHap]

@AjitPai @Qualcomm @JohnEdwardSmee @SampathHemanth @QCOMResearch @QualcommPolicy @john_kuzin As 6G and AI converge, trust infrastructure will become just as important as network infrastructure. One emerging question is how to verify that a human authorized a specific consequential AI action at execution time.

[Dr Harjinder Sandhu @drSandhuHap]

@durga_malladi @Qualcomm @AjitPai @CTIA As the network becomes infrastructure for intelligence, trust has to become infrastructure too.

As agents move from answers to actions, proving human authorization at execution time will become critical.

[Dr Harjinder Sandhu @drSandhuHap]

Authentication proves identity. Permissions define access.

Agentic AI adds another question:

Did the human authorize this specific action at this specific moment?

That is the problem Execution-Time Human Authorization is designed to solve.