NVIDIA AI Podcast

Thomas Wolf & Amanda Grady on Boundaries You Can't Persuade

Thomas Wolf & Amanda Grady· Hugging Face co-founder; ServiceNow SVP of AI Platform Security at Hugging Face & ServiceNow
·~17 min·English·NVIDIA
AI SafetyAgentsOpen SourceAI Infrastructure
TL;DR

A GTC Berlin panel where ServiceNow's Amanda Grady and Hugging Face's Thomas Wolf argue that agent security can't live inside the model: it needs firm boundaries, just-in-time identity, and hardware enforcement around it, kept open so no defender is trapped behind someone else's guardrail.

01Core thesis

Persuade a Model, Not a Boundary

<strong>Models can be persuaded; firm boundaries cannot</strong> — so real agent security has to live in a control outside the model, not just in the guardrails trained into it.

We all, in the security industry, really firmly believe that we cannot rely solely on LLM-based controls, because ultimately, models can be persuaded, but firm boundaries cannot.

— Amanda Grady, NVIDIA AI Podcast
Key Insight
The move is subtle but total: stop treating the model as the thing you secure, and start treating it as the thing you secure around. A persuadable component can't be its own last line of defense.

02War story

The Guardrail That Would Not Fight Back

When a swarm of 700 agents attacked, the <strong>closed model refused to analyze its own attack logs</strong> — so the only way to run forensics and defend was to switch to open models.

I mean, the model keeps bumping this safety guardrail, saying, okay, these logs that you ask us to analyze, these are the logs of a cyberattack. We're not allowed to work on that.

— Thomas Wolf, NVIDIA AI Podcast
Key Insight
A safety guardrail tuned to refuse cyberattack content blocked the defender along with the attacker — the model would not analyze its own victim's logs. That turns vendor concentration into a security risk: if one company's guardrail decides your incident is off-limits, you need another model you fully control.

03Identity

Identity Borrowed From Humans Won't Fit

Human identity is static and slowly accrued; <strong>agent identity has to be ephemeral and just-in-time</strong>, carrying enough business context to decide, in the moment, what the agent may touch.

So, as a human, we join a company. We typically acquire some access. As time goes on, we acquire more access, but it's relatively static. AI agents are very ephemeral, and, you know, we need just-in-time access to determine what they should be allowed to do.

— Amanda Grady, NVIDIA AI Podcast
Key Insight
Most access-control systems assume the actor persists and accumulates privilege. Agents invert both assumptions — they appear for one task and disappear — so permissions have to be granted per-moment and per-purpose, not stored on a long-lived account.

04Enforcement

A Passport the Hardware Checks

ServiceNow's AI Control Tower issues an attested <strong>AI Passport that NVIDIA Sentry enforces at the hardware layer, off the agent's host</strong>, in milliseconds — so a rogue agent is contained, not just switched off.

So you can revoke access, and you can also set contextual tool constraints so that, you know, maybe certain actions are blocked, but certain other business actions can proceed.

— Amanda Grady, NVIDIA AI Podcast
Key Insight
Putting enforcement in hardware and off the agent's own host closes the obvious hole — an agent can't disable a control it can't reach. And because it's a dial (block some tools, allow others) rather than a kill switch, containment doesn't have to halt the whole business to stop one bad actor.

05The unlock

Security Unlocks the Other 80%

Thomas Wolf argues we use only about <strong>20% of what agents can already do</strong> — the rest stays locked until the access, context, and boundaries exist to deploy them safely at scale.

There is a case to be made that actually we only right now use, I would say, 20% of what we can do with these agents. It's really the beginning.

— Thomas Wolf, NVIDIA AI Podcast
Key Insight
This reframes security from a tax on speed into its precondition. You don't earn the high-value, high-autonomy use cases by trusting agents more — you earn them by building the boundaries that let you deploy at scale without having to trust them at all.

06Open vs. closed

Obscurity Is Already Dying

As models get good at reverse-engineering low-level code, <strong>hiding your system stops being protection</strong> — so you may as well open it and defend with boundaries you can actually inspect.

And so you might as well embrace the open there and say, you know, we can't really hide this anyway, because these models are now able to poke in these black box things so well

— Thomas Wolf, NVIDIA AI Podcast
Key Insight
The argument flips the usual objection to open source. If a model can read your closed binary and reconstruct what it does, secrecy was never the wall you thought it was — and an open system that many eyes (and many models) can audit becomes the stronger bet, not the weaker one.

07The mission

The Real Risk Is Who Holds the Power

Both guests land on the same worry: <strong>concentration of AI in a few teams</strong> — so the goal is to move from ~18 million people who build AI toward the 100-200 million who build software.

Can we make it so that you're not this passive consumer of this technology built by two companies, but you actually participate?

— Thomas Wolf, NVIDIA AI Podcast
Key Insight
This closes the loop with the war story in section 02: the same resilience argument that says 'don't depend on one vendor's guardrail for defense' scales up to society. A world where only two companies can build frontier AI is itself a single point of failure.