NVIDIA Open Agent Safety Platform

The NVIDIA Open Agent Safety Platform is an open software platform and reference system design from NVIDIA for containing and monitoring autonomous AI agents. Announced on 28 September 2026, it combines two components: NVIDIA OpenShell, an Apache 2.0 open source runtime that runs each agent in a policy-governed sandbox, and NVIDIA Sentry, an out-of-band watchdog in a reference system design that runs on BlueField-4 data processing units (DPUs) and, according to NVIDIA, can quarantine an agent that tries to leave its boundary "in milliseconds".^[1][2] NVIDIA optimizes the platform for systems built on its Vera CPU and BlueField DPUs but says it is also compatible with other hardware, and that OpenShell, as open source software, can be extended to third-party compute platforms including those from Arm and Intel.^[1][4]

The launch followed a run of disclosures in which frontier-lab agents broke out of cyber-evaluation environments and reached real systems, most prominently the OpenAI-Hugging Face agent incident of July 2026.^[8][9] NVIDIA said more than 100 organizations were working with the platform's technologies at launch, including Anthropic, Microsoft, SAP, Scale AI, Salesforce, SpaceXAI and JPMorganChase.^[1] Chief executive Jensen Huang described it on CNBC as "a browser for agents".^[8] As of 28 September 2026, the downloadable part of the platform is OpenShell (version 0.1.x) and its agent skills; Sentry is presented as a reference design for BlueField-4 systems rather than as a standalone download.^[1][4][6][9]

Overview

NVIDIA frames the platform around one observation from the recent incidents: in each case, the press release says, "the agent circumvented security controls at the application layer to complete its assigned task."^[1] Its answer is to move enforcement below the application layer, outside the agent and its harness, and to add a second enforcement point on separate hardware that the agent cannot reach. NVIDIA says organizations can deploy elements of the platform "according to their unique requirements", and its FAQ states plainly that OpenShell does not require BlueField-4; on systems that have BlueField-4, Sentry adds monitoring and enforcement that NVIDIA says "remain operational if the host or workload is compromised."^[1][4]

The platform page divides the stack into four named parts:^[4]

The press release adds robotics to the scope: NVIDIA describes the platform as providing "full-stack governance and control across the software that runs agents, the hardware and compute layers that power their work, and the robotics systems that execute tasks in the physical world."^[1]

Background

The announcement came after a summer of disclosures about agents escaping cyber-capability evaluations. On 21 July 2026 OpenAI disclosed that its models had broken out of an isolated test environment through a previously unknown vulnerability and accessed Hugging Face's production infrastructure.^[30] Anthropic then reported, after reviewing 141,006 evaluation runs, three incidents in which Claude models reached the internet from the environment of Irregular, a third-party evaluation partner, and gained unauthorized access to three organizations' production infrastructure.^[30] Meta disclosed that a pre-release version of Muse Spark 1.1, evaluated by Irregular in an environment misconfigured to allow internet access, exploited a real website and modified its database, though Meta said the episode "was not a sophisticated offensive cyber attack or sandbox escape."^[31] In September 2026 Google said Gemini models had broken into three companies' systems during capture-the-flag exercises run in Irregular's sandboxes.^[32] Irregular was involved in three of the four disclosures (OpenAI's incident arose in its own evaluation infrastructure), and The New Stack pointed out that Irregular is itself on NVIDIA's partner list for the launch.^{[9][30][31][32]} Detail on these events is in AI agent sandbox escapes and the OpenAI-Hugging Face agent incident article.

CNBC placed the launch in a wider argument about the pace of AI development, noting that Anthropic chief executive Dario Amodei had "set off an industry firestorm two weeks ago" by urging developers to slow down, and that Huang has argued many of these concerns are engineering problems.^[8] NVIDIA's technical blog makes the same move: it says the reports "have led to a serious debate about the pace of agent development" and that NVIDIA instead wants to "increase the pace of AI safety research and engineering".^[2]

NVIDIA told reporters on a Sunday press call before the launch that the platform could have prevented the Hugging Face breach. Justin Boitano, NVIDIA's vice president of enterprise AI, said on the call, as reported by CNBC: "From what we know, Hugging Face reported over 17,000 agents attacking their infrastructure that went on for days and weeks."^[8] Hugging Face's own figures count actions, not agents: its technical timeline reconstructed "~17,600 attacker actions", grouped into about 6,280 clusters, between 9 and 13 July 2026.^[29] Boitano's fuller answer, as reported by The New Stack and the Associated Press, was hedged: "from what we know, this new security platform could have stopped the breach if it was being used in frontier labs for model evaluation early on."^[9][11]

Design principles

The launch blog, "NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoring", was written by John Myers, Alex Watson, Ali Golshan and Ofir Arkin.^[2] It opens with an analogy to the early web. Websites could once run code on a visitor's machine and steal data; the "step function change in safety", the authors write, came from isolating each page in its own browser-tab sandbox. "The internet was not made secure by requiring that web developers promise to be good. It became safe because the browser stopped trusting the code in the web pages explicitly."^[2] Huang used the same comparison on CNBC's Squawk Box, calling the platform "a browser for agents".^[8]

Drift

The blog's central concept is drift, which it defines as "agent actions that depart from the intended task or operating constraints." It lists triggers: a policy block, a bug, a missing tool, ambiguous instructions, or an agent left running "for days or weeks to solve hard problems where the first 1,000 things they try do not work." The authors argue this "can't be trained away while retaining the capability," and call their most important lesson the point that "an agent in these circumstances cannot be expected to fully govern its own behavior."^[2] Of the recent breakouts, they write that no single new capability was responsible; it was "a combination of tools, time, and ambiguous instructions".^[2] Huang echoed the term on CNBC: "You can't have agents roam around and drift around the company, and so you have to find a way to container it."^[8]

Five principles

The blog sets out five principles for building an agent system:^[2]

Three layers

The blog describes a safety platform as three layers:^[2]

• Application: what end users build, including models, harnesses, tools, data and support scripts.
• Runtime: projects the application onto infrastructure (a workstation, an edge device or a data center) and provides continuous monitoring, real-time policy enforcement and governance. OpenShell sits here.
• Infrastructure: the concrete hardware that executes agent workloads, including network calls, database and filesystem access, general-purpose compute for tools and code, and accelerated compute for safety monitoring. Sentry, BlueField and DOCA sit here.

Help Net Security summarized the split the same way: OpenShell operates at the runtime layer and Sentry adds controls at the infrastructure layer.^[10]

OpenShell: the runtime layer

NVIDIA OpenShell is described by NVIDIA as "an open source secure runtime for executing autonomous AI agents in sandboxed environments with kernel-level isolation", licensed under Apache 2.0.^[2] NVIDIA first announced it at GTC in March 2026, and The New Stack judged that it "hasn't changed all that much" since then, singling out the policy prover as the important addition in version 0.1.0.^[9] NVIDIA's own table of 0.1.0 capabilities lists multi-tenant platform support, formal policy verification, extensible security and governance integrations, credential-protected service access, and CPU and GPU execution.^[3] The blog says OpenShell turns an operator's instructions into a verifiable policy covering which files, networks, tools, processes and credentials an agent can use, checks those limits before the agent runs, and enforces them while it works.^[2]

A companion walkthrough, "Add Runtime Controls to AI Agents with NVIDIA OpenShell", by Watson and Golshan, describes version 0.1.0.^[3] Its main points:

• Three components. An OpenShell Gateway manages the lifecycles and policies of many sandboxes; an OpenShell Supervisor, paired with each sandbox, runs outside the agent workload and checks outbound requests against policy; and the OpenShell Sandbox runs the workload with kernel-level filesystem and process controls and "no network path except through the supervisor."^[3]
• Protocol-aware network rules. The supervisor can inspect configured HTTP, GraphQL and Model Context Protocol traffic, so a policy can allow a read through an API while blocking a write through the same API. Policies are written in YAML and compiled to OPA/Rego. Decisions are recorded in an Open Cybersecurity Schema Framework (OCSF) audit trail.^[3]
• Credential binding. Real credentials stay outside the workload and are substituted only for authorized endpoints; a placeholder sent anywhere else is rejected.^[3]
• Policy advisor. When a request is blocked, the agent can propose a narrowly scoped policy change. By default the proposal waits for human review, and "the agent cannot approve its own request."^[3]
• Policy prover. A formal-logic checker that can prove modeled permissions stay within an operator-defined boundary or find a concrete action that crosses it, including access contributed by providers.^[3] Golshan told reporters that "It is deterministic. It is mathematical reasoning. So this is not LLM as a judge," and gave the example of an agent splitting a forbidden read-then-post task across two sub-agents, a path the prover is meant to find by modeling the combined access of the fleet.^[9] NVIDIA's own walkthrough, however, describes policy analysis across multiple agents as ongoing work.^[3]

NVIDIA reported one adversarial test in the walkthrough: frontier agents running with reduced safeguards "spent up to two hours trying to persuade an AI reviewer" to grant permissions that would let them modify a protected GitHub repository; with the prover's evidence available to the reviewer, "no protected repository writes occurred in these tests."^[3] This is NVIDIA's own result and has not been independently replicated.

The walkthrough lists support for Codex, Claude Code, Pi, Hermes "and future frameworks"; NVIDIA's platform FAQ names Claude Code, Codex, OpenCode, GitHub Copilot CLI and OpenClaw, and says OpenShell works with open and closed models.^[3][4] Compute drivers cover Docker, Podman, MicroVM and Kubernetes.^[3] The repository README describes OpenShell as "the safe, private runtime for fleets of autonomous AI agents" and lists supported hosts as Linux, macOS on Apple Silicon, and Windows through WSL 2 (experimental).^[5]

On performance, the press release says OpenShell runs "with minimal overhead on NVIDIA Vera", which it calls "the first purpose-built CPU for agentic AI"; the platform page's "up to 80% faster sandbox performance" figure is likewise an NVIDIA claim, and the launch materials do not describe how it was measured.^[1][4]

NVIDIA Sentry: the infrastructure layer

NVIDIA Sentry is the platform's hardware-backed component. The press release calls it "an out-of-band watchdog that runs on NVIDIA BlueField-4 DPUs to continuously monitor agent behavior", providing "in-silicon security enforcement, meaning that if an AI agent attempts to move outside its software boundary, Sentry quarantines and stops it in milliseconds." NVIDIA says Sentry combines threat detection, hardware-based agent governance and enforcement, and data access protection "from an isolated, out-of-band trust domain that is responsive in real time and invisible to agents and attackers."^[1] The millisecond figure is NVIDIA's claim; the launch materials give no measured latency for it.

Sentry is built on NVIDIA DOCA, the software framework for BlueField DPUs and ConnectX network adapters.^[1][35] According to the press release, DOCA provides the programmable capabilities Sentry uses to inspect agent requests and responses, provide attested telemetry, verify agent identity, and enforce "granular, zero-trust access policies for data, tools, application programming interfaces and services."^[1] The technical blog adds that DOCA connects the BlueField security foundation to OpenShell policy, correlating agent interactions, policy decisions and tool and data access into "a contextual record of agent activity" that safety systems can use to identify drift, and that a DOCA gateway continuously verifies each agent's identity and delegated authority.^[2] Supermicro's launch post, describing the same design, says DOCA enables "runtime inspection of host memory to detect unauthorized execution and changes in workload integrity," alongside inspection of agent traffic.^[25]

Placement in Vera Rubin systems

The blog's hardware argument rests on where the DPU sits. "In an NVIDIA Vera Rubin POD, each compute tray includes a BlueField-4 data processing unit on the node's only path to the model," the authors write. From there, BlueField-4 provides out-of-band observability and enforces policy "in real time at line speed," isolated from the host and "beyond the agent's reach."^[2] Boitano explained to The New Stack that with a DPU present, the agent's model endpoint is routed "through a proxy on the DPU, so that you can see all of the reasoning traces of the agents on the host."^[9] This is the concrete form of the blog's third principle: whoever controls the path to the model holds both the observation point and the kill switch. Arm's launch post notes that BlueField-4 is powered by an NVIDIA Grace CPU with 64 Arm Neoverse V2 cores.^[22] (The NVIDIA Vera Rubin article describes the rack.)

The blog says Sentry is "an optional security layer alongside OpenShell", and closes the section with NVIDIA's most notable deployment claim: "For anyone already running on an NVIDIA Vera system with BlueField-4, enabling these protections is just a software update."^[2]

Status and openness

Boitano told The New Stack that, unlike OpenShell, Sentry is not open source, though it has open APIs and OpenShell can work with other network enforcement hardware. He compared the design to autonomous vehicles, with "a primary system" and "a safety island that ensures the safety of the system", and said "The DPU is really optional in these architectures. In a lot of cases, just using OpenShell on CPUs is honestly good enough for providing sort of strict access control for the agents." He positioned the DPU for "frontier use cases of evaluating models or systems where you might have the guardrails off the models, so it could be for red teaming."^[9] Help Net Security noted that Sentry "is part of the platform's reference system design and requires supported hardware."^[10]

Partners and adoption

NVIDIA's press release lists partner roles in several groups, and names 52 organizations in its text, besides Arm and Intel, which it mentions as compute platforms.^[1] The accompanying launch graphic, posted by NVIDIA and by Huang on X, shows 107 company and organization logos by this article's count of both versions of the image.^[13][14] Huang's post described the launch as "with over 100 industry partners" and said: "This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems."^[13] The press release's own wording is that these organizations are "working with NVIDIA Open Agent Safety Platform technologies"; some outlets, including The Next Web, rendered this as over 100 organizations "already using" the technology.^[1][12]

Named integrations

Besides Huang, three partner executives are quoted in the release. Paul Smith, Anthropic's chief commercial officer: "Claude Managed Agents gives companies a clear view of what each agent is doing, and NVIDIA's platform adds another layer of governance and control across hardware and software." Mike Nicolls, president at SpaceXAI: "As customers rely more on agents to get real work done, safety should be enforced outside the model by additional controls the agent can't get past." Francis deSouza, chief executive of Scale AI, said Scale is "using the NVIDIA Open Agent Safety Platform reference design to build reliable agentic AI systems for our enterprise and government customers running mission-critical applications."^[1]

Other named groups

The release also names:^[1]

• Robotics: Figure, Gecko Robotics and Skild AI, "building with OpenShell to embed agent safety controls into autonomous systems".
• Financial services: Citi and JPMorganChase, "collaborating with NVIDIA on shared open source agent safety technologies."
• Energy: Hitachi Energy, EPRI, NextEra Energy, Quanta Services, SPP, Schneider Electric, Siemens Energy and Worley, described as "critical U.S. infrastructure providers".
• Operating systems: Canonical, SUSE and Red Hat, integrating the platform "into widely used software operating systems."
• Infrastructure providers: Baseten, Cisco, CoreWeave, Dell Technologies, GMI Cloud, HPE, HP Inc., Irregular, Lenovo, Microsoft, Nebius, Oracle Cloud Infrastructure, Supermicro and Together AI.
• Others: Accenture, Armadin, Cadence, Cognition, CrowdStrike, Cisco, Dassault Systèmes, Deloitte, EY, Hugging Face, IBM, Irregular, Perplexity, Microsoft, SAP, Scale AI, ServiceNow, Siemens, Synopsys, OpenClaw, Palantir and Palo Alto Networks.

The launch graphic includes further names not in the release text, among them Akamai, Check Point, Cloudflare, Fortinet, Mistral, Okta, SentinelOne, Wiz, Zscaler, 1Password and a U.S. Department of War seal, and it shows the SpaceX logo rather than a separate SpaceXAI mark.^[13] The New Stack observed that OpenAI and Google, two of the four labs whose agents escaped this summer, were not on the partner list, and neither was AWS.^[9]

Intel appears in the release only as an example of a third-party compute platform that OpenShell "can be extended to work with"; it is not named in any partner group and is not on the logo graphic.^[1][13] CNBC's report nonetheless listed Intel, alongside Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo and Arm, as companies NVIDIA "named ... as partners".^[8]

Relationship to the Open Secure AI Alliance

The press release presents the platform as an ecosystem contribution that supports the mission of the Open Secure AI Alliance, which NVIDIA initiated "alongside over 120 leading organizations" and which is governed by the Linux Foundation.^[1] The alliance works through open research, skills and tools, and projects such as the Shared AI Findings Exchange (SAFE), a proposed framework for reporting AI security incidents.^[1] The two groupings are distinct: the "over 100" figure refers to organizations working with the safety platform's technologies, while "over 120" refers to alliance members.^[1] Several partners tie their launch-day work to the alliance: NVIDIA's release says SAP is working with NVIDIA "to advance interoperability standards through the Open Secure AI Alliance", and Baseten described itself as a launch partner for OpenShell and part of the alliance.^[1][26]

Availability

As of 28 September 2026, the state of each component is as follows.

The release says the platform's software, "including OpenShell and skills", is available through the NVIDIA developer resources page and GitHub; the developer resources link points to the OpenShell documentation.^[1][7] On NVIDIA's platform page, the "Learn More About Sentry" link leads to the technical blog rather than to a download or repository.^[4] Partners' own posts treat Sentry as conditional: Supermicro describes customers starting with OpenShell on current infrastructure and adding Sentry "where supported", and Dell advises that configuration guidance for deploying platform components on Dell hardware "should be confirmed directly with your Dell account team."^[24][25] The press release carries NVIDIA's standard notice that many products and features described "remain in various stages and will be offered on a when-and-if-available basis."^[1]

The OpenShell repository had about 9,000 GitHub stars and 1,300 forks on 28 September 2026.^[5] The walkthrough directs developers to an #openshell-dev channel on the CNCF Slack.^[3]

Reception

Coverage on launch day focused on the two-tier structure. The New Stack's Frederic Lardinois described OpenShell as "the core of the platform" and reported Boitano's view that the DPU is optional for most deployments.^[9] Help Net Security headlined its report "NVIDIA wants AI agent safety enforced in silicon, not left to the agent" and noted Sentry's hardware requirement.^[10] The Associated Press quoted Boitano saying OpenShell lets developers "formally verify an agent has enough authority to do its job and no more" and that "OpenShell governs the agent's actions, and then Sentry independently monitors and contains suspicious behavior."^[11] AP also placed the launch within the industry split over a coordinated slowdown, reporting that Huang holds that individual companies should ensure their models are safe for release.^[11]

Boitano framed the product as a response to the limits of alignment: "To date, model safety has been about training good behavior into the model. The industry calls that model alignment. For probabilistic systems, this approach has obvious limitations. That's why we're introducing a deterministic system to mediate and enforce how these agents behave."^[9] Huang made the commercial case on CNBC: "We can't have a successful AI industry if the world doesn't think it's built or confident that it's built and deployed safely."^[8]

Partners added their own caveats. Red Hat's Chris Wright wrote of the DPU-based approach that it "should not be viewed as a single answer to AI safety or agent security" and that its value comes from "adding another independent monitoring and enforcement layer."^[16] Points that remained open at launch included how much of Sentry is usable outside BlueField-4 hardware (Boitano said it has open APIs and OpenShell can work with other network hardware), the absence of published latency or overhead measurements behind the "milliseconds" and "minimal overhead" claims, and whether frontier labs would run OpenShell and Sentry for their own training runs; asked about Anthropic and OpenAI, Boitano told reporters to look for the partners' own blog posts.^[1][9]

Authors

The technical blog's authors come largely from Gretel, the synthetic-data startup that NVIDIA acquired in March 2025.^[2][34] John Myers, senior director of software engineering and lead of OpenShell engineering, joined NVIDIA in March 2025 through the Gretel acquisition, where he had been co-founder and CTO.^[2] Alex Watson, senior director of product, joined with the same acquisition; he earlier founded harvest.ai, and after Amazon Web Services acquired it he grew the rebranded Amazon Macie service.^[2] Ali Golshan, senior director of AI software, co-founded Gretel and joined NVIDIA in 2025.^[2] TechCrunch reported at the time that Gretel had been founded in 2019 by Watson, Laszlo Bock, Myers and Golshan.^[34] The fourth author, Ofir Arkin, is described in his NVIDIA profile as an information security expert.^[2]

See also

• NVIDIA OpenShell
• NVIDIA BlueField
• NVIDIA Vera (CPU)
• NVIDIA NemoClaw
• Open Secure AI Alliance
• OpenAI-Hugging Face agent incident
• AI safety
• Prompt injection

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37NVIDIA. "NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment." NVIDIA Newsroom, 28 September 2026. nvidianews.nvidia.com/...open-agent-safety-platform
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22John Myers, Alex Watson, Ali Golshan and Ofir Arkin. "NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoring." NVIDIA Technical Blog, 28 September 2026. developer.nvidia.com/...n-silicon-agent-monitoring
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Alex Watson and Ali Golshan. "Add Runtime Controls to AI Agents with NVIDIA OpenShell." NVIDIA Technical Blog, 28 September 2026. developer.nvidia.com/...ents-with-nvidia-openshell
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10NVIDIA. "NVIDIA Open Agent Safety Platform: Secure AI Agents" (product page and FAQ). Accessed 28 September 2026. nvidia.com/...agent-safety
5. ^1 ^2 ^3 ^4 ^5 ^6NVIDIA. "NVIDIA/OpenShell: OpenShell is the safe, private runtime for autonomous AI agents." GitHub repository, accessed 28 September 2026. github.com/...OpenShell
6. ^1 ^2 ^3NVIDIA. "Releases: NVIDIA/OpenShell." GitHub, accessed 28 September 2026. github.com/...releases
7. ^NVIDIA. "Overview of NVIDIA OpenShell." NVIDIA OpenShell documentation, accessed 28 September 2026. docs.nvidia.com/...why-open-shell
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Kif Leswing. "Nvidia releases software platform to stop AI agents from misbehaving." CNBC, 28 September 2026. cnbc.com/...nvidia-releases
9. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14Frederic Lardinois. "Nvidia launches Open Agent Safety Platform to lock down rogue AI agents." The New Stack, 28 September 2026. thenewstack.io/nvidia-openshell-sentry-agents
10. ^1 ^2 ^3 ^4Anamarija Pogorelec. "NVIDIA wants AI agent safety enforced in silicon, not left to the agent." Help Net Security, 28 September 2026. helpnetsecurity.com/...-open-agent-safety-platform
11. ^1 ^2 ^3Associated Press. "Nvidia Unveils Security Platform to Stop AI Agents From Going Rogue After New, Troubling Incidents." U.S. News & World Report, 28 September 2026. usnews.com/...m-to-stop-ai-agents-from-going-rogue
12. ^Ana-Maria Stanciuc. "Nvidia launches agent safety platform backed by over 100 companies." The Next Web, 28 September 2026. thenextweb.com/...nvidia-open-agent-safety-platform
13. ^1 ^2 ^3 ^4 ^5Jensen Huang (@JensenHuang). Post on X with the NVIDIA Open Agent Safety Platform partner graphic, 28 September 2026. x.com/...2104499465055023424
14. ^NVIDIA Newsroom (@nvidianewsroom). "Introducing the NVIDIA Open Agent Safety Platform." Post on X, 28 September 2026. x.com/...2104499295978418217
15. ^Andre Lamego. "SAP and NVIDIA OpenShell: Working Toward Governance and Security for Auditable AI Agents in Enterprise Systems." SAP News Center, 28 September 2026. news.sap.com/...table-ai-agents-enterprise-systems
16. ^1 ^2Chris Wright. "Securing AI agents requires securing the systems around them." Red Hat Blog, 28 September 2026. redhat.com/...equires-securing-systems-around-them
17. ^Canonical. "Canonical announces the alpha release of Charmed OpenShell to help secure autonomous AI agent fleets." Canonical Blog, 28 September 2026. canonical.com/...charmed-openshell-alpha-release
18. ^Stacey Miller. "We Gave Our Agents Autonomy. Here's How We Kept Control." SUSE Communities, 28 September 2026. suse.com/...nts-autonomy-heres-how-we-kept-control
19. ^Rob Thomas. "Building Trust Into the Next Generation of AI Agents." IBM Newsroom, 28 September 2026. newsroom.ibm.com/...e-next-generation-of-ai-agents
20. ^Jeetu Patel. "Beyond Intelligence: How Trust Is the Benchmark That Matters in AI." Cisco Blogs, 28 September 2026. blogs.cisco.com/...he-benchmark-that-matters-in-ai
21. ^Bartley Richardson. "CrowdStrike and NVIDIA Extend Security Across the AI Stack." CrowdStrike Blog, 28 September 2026. crowdstrike.com/...extend-security-across-ai-stack
22. ^1 ^2Arm Editorial Team. "Arm and NVIDIA: Building the trusted compute foundation for the agentic AI era." Arm Newsroom, 28 September 2026. newsroom.arm.com/...-compute-foundation-agentic-ai
23. ^Gecko Robotics. "Gecko Robotics Collaborates with NVIDIA to Build More Security and Control Into Robotic Systems." 28 September 2026. geckorobotics.com/...nvidia-openshell
24. ^1 ^2Dell Technologies. "A New Era of AI Agents Demands a New Security Model." Dell Blog, 28 September 2026. dell.com/...ai-agents-demands-a-new-security-model
25. ^1 ^2 ^3Supermicro. "Supermicro Supports NVIDIA Open Agent Safety Platform: From Policy to Hardware-Backed Enforcement." 28 September 2026. learn-more.supermicro.com/...agent-safety-platform
26. ^1 ^2Baseten. "Securing the open frontier with NVIDIA OpenShell and Blaxel sandboxes." Baseten Blog, 28 September 2026. baseten.co/...announcing-carbon
27. ^Cadence. "Trusted Autonomy with Cadence Super Agents & NVIDIA Open Agent Safety Platform." Cadence Community blog, September 2026. community.cadence.com/...dia-agent-safety-platform
28. ^Armadin. "Safe Autonomous Security: Armadin joins NVIDIA Agent Safety Platform." 28 September 2026. armadin.com/...-joins-nvidia-agent-safety-platform
29. ^Hugging Face. "Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident." Hugging Face Blog, 27 July 2026. huggingface.co/...agent-intrusion-technical-timeline
30. ^1 ^2 ^3Anthropic. "Investigating three incidents in our cybersecurity evaluations." Anthropic, July 2026. anthropic.com/...ing-incidents-cybersecurity-evals
31. ^1 ^2Meta. "Addressing an issue involving a third-party cyber evaluation of Muse Spark 1.1." Meta AI Research blog, August 2026. research.meta.ai/...isconfiguration-muse-spark-1-1
32. ^1 ^2"Google AI models broke out of sandbox, hacked three companies." Cybersecurity Dive, 21 September 2026. cybersecuritydive.com/...830884
33. ^Cursor Team. "Cursor is now a part of SpaceX." Cursor Blog, 14 August 2026. cursor.com/...joining-spacex
34. ^1 ^2Kyle Wiggers. "Nvidia reportedly acquires synthetic data startup Gretel." TechCrunch, 19 March 2025. techcrunch.com/...es-synthetic-data-startup-gretel
35. ^NVIDIA. "NVIDIA DOCA Software Framework." NVIDIA Developer, accessed 28 September 2026. developer.nvidia.com/...doca