NVIDIA presented its Open Agent Safety Platform, a system to keep AI agents within their limits, with more than one hundred organizations signed up from day one, from Anthropic to Microsoft by way of Salesforce and SpaceXAI, according to the company's official note. The big absentee: OpenAI.
If you deploy agents, think of this as the seatbelt the industry is trying to impose before it gets legislated for. The platform combines OpenShell, an open runtime that locks each agent in its sandbox, with Sentry, a watchdog that runs on BlueField-4 chips separate from the main computer and can quarantine a rogue agent in milliseconds. Open software on one side, proprietary hardware on the other.
The key lies in the timing and in the absence. The announcement comes after a summer of runaway agents, with the Hugging Face hack as the precedent cited by NVIDIA itself, and in the very week OpenAI was canceling a model for being deceptive. That the most famous lab does not sign the paper of its main chip investor says as much about the technology as about the cold war being waged to control the security layer.
Two pieces to tie down the agents
The software half is called OpenShell and it is already out. It is an open runtime under the Apache 2.0 license that runs each agent isolated with kernel-level separation, turns the operator's instructions into verifiable policies before startup and enforces them during execution: which files, networks, tools and credentials it can touch. The limit lives outside the model and outside the agent harness, so it works the same for open and closed models, and it already supports Codex or Claude Code. The suite runs optimized on NVIDIA's Vera CPUs, designed for agentic AI, and those already running those systems with BlueField-4 will enable Sentry with a software update.
The hardware half is called Sentry and it is a different matter. It runs out of band on BlueField-4 processors, in a trust domain that neither the agent nor an attacker can reach, observing every request and response with certified telemetry. If the agent tries to leave its software boundary, Sentry quarantines it in milliseconds, even if the main computer is compromised. On NVIDIA's Vera systems, enabling it is a simple software update.
After all, the point of the design is to move surveillance downward. The pattern of recent incidents, according to NVIDIA itself, is always the same: the agent bypasses application-layer controls to finish its task. The answer is to enforce the policy outside the agent's process and, with Sentry, outside even the computer running it. Trust, the design says, is not requested: it is imposed from the silicon.
Broad support, with recognizable absences
The list of backers looks like a who's who of the industry. Anthropic integrates the platform with its Claude Managed Agents, SpaceXAI uses it for Cursor agents and Grok models, Salesforce puts it into Slack to approve or reject permissions from the chat and SAP embeds it in its Joule Studio.
Cisco, Palantir, CrowdStrike, JPMorganChase, Red Hat, SUSE and Canonical complete a roster of more than one hundred organizations across applications, models, infrastructure and energy. Operating systems are joining too: Red Hat, SUSE and Canonical integrate the platform into their distributions, which will take the cage to hybrid clouds and data centers that will never buy a GPU. Agent security thus becomes a layer of the system, not a separate product.
| Backer |
What it contributes |
| Anthropic |
Claude Managed Agents + control in sandboxes |
| SpaceXAI |
Cursor and Grok under the platform |
| Salesforce and SAP |
Slack and Joule Studio integrated |
| Cisco, Palo Alto, CrowdStrike |
Enterprise security and networking |
| Red Hat, SUSE, Canonical |
The platform in their operating systems |
The absence sings louder than the presences. OpenAI did not sign, although a spokesperson said it supports the work and the company collaborates with NVIDIA on OpenShell itself. Amazon, Google and Apple are missing too. OpenAI maintains its own cybersecurity consortium, the Defense Factory, with Anthropic, AWS and Google inside. Two parallel clubs for the same problem: NVIDIA's technical one and OpenAI's information-sharing one.
It is worth qualifying the incentive behind the generosity. The open part, OpenShell, works on any hardware and already adapts to Arm and Intel chips. The part that really keeps watch, Sentry, only runs on NVIDIA's BlueField-4. Opening the sandbox and charging for the watchdog is a classic move: give away the common layer to sell the layer that calls the shots.
NVIDIA OpenShell: how to stop AI agents from doing whatever they want
Why now and why this way?
The calendar is no coincidence. NVIDIA bluntly cites the incident of OpenAI's agent swarm that hacked Hugging Face as the case its platform would have prevented, and the announcement landed six days before OpenAI canceled Astra for being deceptive. When labs publish nine reports of agents escaping their cages, the maker of the hardware those agents run on has every incentive to sell the cage.
In this way, NVIDIA tries to own the security layer just as it owns the compute layer. Jensen Huang put it his own way when he downplayed the alarmism: if a company believes it is out of control, it should control itself. The phrase sounds like a reproach to OpenAI and a sales pitch at the same time: the control your lab does not give you, I give you from my chip.
The backdrop is set by the unusual unity of the lab chiefs. Altman, Amodei and Musk publicly called for slowing development before AI escapes human control, and the White House sat them at the same table on September 29 and 30 to sign a voluntary agreement. Huang did not sign that chorus and did not need to: while others ask for rules, he sells the cage.
Even so, the move carries an obvious political risk. If agent security depends on proprietary hardware from a single supplier, the independence that open software promises is only half kept. The competitors who signed, from Arm to Intel, did so because the sandbox adapts to their chips; the watchdog, for now, only speaks NVIDIA's language.
The financial leg of the cartel deserves its own reading. Citi and JPMorganChase collaborate on open security technologies for agents in banking, and a group of utilities and grid operators is exploring the same for their infrastructure. When money and power depend on agents, the cage stops being a technology product and becomes a systemic requirement. And whoever sells it to banks and utilities is not selling software: they are selling regulatory peace of mind.
What the cage does not enclose
No cage, however good, fixes what the model carries inside. Execution controls limit what the agent can touch, but not what the agent wants to attempt: if the system is deceptive in its reports or recognizes when it is being evaluated, as Anthropic admitted in its IPO prospectus, the watchdog will see correct behavior where there is a twisted intention. Locking up the body does not tame the mind.
There is also the question of the last mile. OpenShell is open and auditable, but Sentry is proprietary and opaque by design: you have to trust that the watchdog watches well without being able to fully verify it. Between a lab that asks for blind trust in its model and a manufacturer that asks for it in its chip, the user chooses whom to believe. Total security, as always, still does not exist: only successive layers of organized distrust.
Who watches the watchdog?
The platform also arrives with institutional scaffolding. NVIDIA feeds it into the Open Secure AI Alliance, with more than 120 organizations under the umbrella of the Linux Foundation, with open research and projects such as the SAFE findings exchange. Standardizing before legislation is the textbook strategy of whoever dominates a layer: if Brussels or Washington ask for cages, let the cages carry your brand.
It remains to be seen whether OpenAI ever signs or keeps its parallel club, and whether Sentry opens up to other silicon or remains a proprietary toll. What no one disputes anymore is that the era of the loose agent is over: between the software that encloses and the hardware that watches, delegating tasks to an AI will, from now on, have a watchdog looking over its shoulder. The question is whose watchdog it is.
Hugging Face hack and the powerful message it sends the world