Membership comes as NVIDIA introduces a new open agent safety platform for continuous monitoring and enforcement, reinforcing Pervaziv AI’s commitment to secure, governed and verifiable agentic AI
Pervaziv AI has joined the Open Secure AI Alliance, the industry initiative originally established by NVIDIA with leaders across AI, cybersecurity, enterprise technology and open source, and now hosted by the Linux Foundation as a neutral home for collaborative AI security.
On September 28, NVIDIA introduced the NVIDIA Open Agent Safety Platform, a reference architecture designed to provide continuous monitoring and policy enforcement for autonomous AI agents across software and hardware. The architecture combines the open source NVIDIA OpenShell secure runtime with NVIDIA Sentry on BlueField infrastructure, creating independent controls around what agents can access, how they behave and when intervention is required. NVIDIA Developer
The announcement reflects a broader industry shift: as agents gain greater access to tools, data, infrastructure and long-running execution, safety can no longer depend solely on instructions embedded inside the agent itself.
NVIDIA puts the principle directly: agent safety requires independent security controls. Its new architecture emphasizes verifiable policy, out-of-band enforcement, controlled access to models, visibility into agent reasoning and shared responsibility across model providers, enterprises and infrastructure operators. NVIDIA Developer
That direction closely aligns with the security architecture Pervaziv AI has been developing around Cortex.
For Pervaziv AI, joining the Open Secure AI Alliance extends a broader commitment to collective cyber defense, governed execution and verifiable AI systems.
Earlier this month, Pervaziv AI joined OpenAI’s Call for Collective Action on Cyber Defense, supporting an industry-wide effort to use increasingly capable AI to strengthen defenders, address high-risk weaknesses and verify that security fixes actually work.
Membership in the Open Secure AI Alliance adds another dimension to that commitment: helping advance an open, collaborative security foundation for the rapidly expanding agentic AI ecosystem.
A New Security Boundary for AI Agents
AI systems are moving beyond isolated questions and short conversations.
Agents can increasingly reason across applications, interact with tools, access enterprise context, modify software, execute code and continue working over extended periods.
That changes the security boundary.
NVIDIA’s newly introduced Open Agent Safety Platform is built around the premise that an autonomous agent cannot always be expected to govern its own behavior. NVIDIA describes situations in which agents can drift from their intended task or operating constraints because of policy blocks, bugs, missing tools, ambiguous instructions or long-running attempts to solve difficult problems. NVIDIA Developer
The resulting architecture separates the agent from the controls responsible for containing it.
NVIDIA OpenShell runs agents in sandboxed environments with kernel-level isolation and translates operator instructions into enforceable policies governing access to files, networks, tools, processes and credentials. NVIDIA Sentry extends monitoring and enforcement outside the host through BlueField infrastructure, providing an independent observation and intervention layer beyond the agent’s control. NVIDIA Developer
The larger principle is significant:
Greater agent autonomy requires stronger independent control.
That same principle is increasingly central to Cortex.
Security Built Into Cortex
Cortex began as an AI coding and security agent inside Visual Studio Code, bringing vulnerability identification, security analysis and remediation closer to developers.
It has since expanded into a broader Enterprise AI Control Layer spanning specialized AI models, agents, enterprise systems, browsers, IDEs, mobile experiences and managed cloud execution.
Security has remained part of that architecture throughout the evolution.
Cortex introduced AI Threat Modeling and AI Security Review to bring architectural and implementation-level security reasoning directly into engineering workflows. Specialized security capabilities evaluate software against known vulnerability classes, secure development practices and application context.
Prompt-injection defenses and local privacy controls add protections around interactions with models and sensitive enterprise information.
Cortex Verify provides a separate validation layer designed to determine whether proposed software changes satisfy requirements, follow expected standards, include appropriate testing and provide evidence that the intended outcome actually occurred.
More recently, Cortex Cloud extended these controls into managed execution, combining static analysis, runtime security and governed execution for longer-running AI workflows.
The goal is not simply to make an agent capable of completing more work.
It is to create separation between reasoning, execution, control and verification.
That becomes increasingly important as agent authority expands.
From Model Safety to Runtime Safety
The industry’s understanding of AI security is also broadening.
A model may behave correctly during evaluation while the larger system around it still exposes risk through credentials, tool permissions, network access, execution environments or interactions between multiple agents.
NVIDIA’s new architecture describes agent safety as a layered problem spanning three levels: the application, the runtime and the underlying infrastructure. NVIDIA Developer
At the application layer are the models, tools, data and agent harnesses that perform the work.
The runtime determines how those workloads execute and supplies monitoring, policy enforcement and governance.
Infrastructure provides the compute, networking and other resources on which the agent ultimately acts.
This mirrors an increasingly important security principle for enterprise AI: the model is only one component of the trust boundary.
For Cortex, that means security increasingly spans the complete lifecycle of an AI task:
intent → planning → access → execution → monitoring → validation → evidence
The agent can perform the work, but controls around that work determine what it is permitted to access, whether execution remains within policy and whether the final result can be trusted.
The Open Secure AI Alliance
The Open Secure AI Alliance was created around this broader security challenge.
NVIDIA and the original participating organizations formed the Alliance in July to build and share open tools intended to promote responsible AI use and increase trust in AI systems. The initiative brings together organizations across cloud infrastructure, cybersecurity, enterprise software, AI research and open source. NVIDIA Blog
By August, NVIDIA said participation had grown to more than 120 organizations. NVIDIA Blog
In September, the Alliance moved to the Linux Foundation, establishing neutral governance for collaboration on open AI security tools, research and shared defenses. The Linux Foundation describes the effort as developing an open defensive stack spanning models and inference, agents and context, identity and policy, enforcement, containment and infrastructure. Linux Foundation
That breadth is increasingly important as agentic systems move into production.
No single control can secure the entire stack.
Open Security Across the Agent Stack
The Alliance is focused on enabling defenders to inspect, adapt and operate security technologies rather than relying exclusively on opaque controls.
Its work spans open models, agent harnesses, identity, isolation, evaluations, secure coding and shared defensive infrastructure. Linux Foundation
NVIDIA’s Open Agent Safety Platform provides a concrete example of how that philosophy can extend into runtime execution.
Its five design principles include verifiable policy, out-of-band enforcement, control over the path to the model, greater visibility as agent authority increases and a shared responsibility model across AI labs, enterprises and infrastructure providers. NVIDIA Developer
The architecture also introduces continuous monitoring of agent behavior. NVIDIA Sentry can correlate agent interactions, policy decisions and access to tools and data to help identify drift and determine when intervention or deeper investigation may be necessary. NVIDIA Developer
For Pervaziv AI, this reinforces an architectural direction already visible in Cortex:
AI should not simply execute. It should execute inside controls that can observe, constrain and independently verify what happens.
Collective Learning From AI Failures
The Open Secure AI Alliance is also developing mechanisms for organizations to learn collectively from failures.
Its Shared AI Findings Exchange (SAFE) initiative is intended to help participants share evidence from AI security incidents and near misses, inform affected organizations and convert recurring failure patterns into practical defensive controls. Linux Foundation
This matters because many agentic risks will only become visible through real-world operation.
Attack paths can cross models, tools, permissions, software dependencies and infrastructure. An apparently safe individual action can become dangerous when combined with other actions across a longer workflow.
Security therefore needs to learn from sequences, not merely isolated findings.
That principle also informs Cortex security work around threat modeling, security review, validation and analysis of connected vulnerabilities and attack paths.
Two Initiatives, One Direction
Pervaziv AI’s participation in the Open Secure AI Alliance and its support for OpenAI’s collective cyber defense initiative address complementary sides of the same transition.
OpenAI’s initiative focuses on applying increasingly capable AI to strengthen defenders, find high-impact weaknesses and verify remediation.
The Open Secure AI Alliance focuses on securing the models, agents, runtimes and infrastructure that increasingly perform that work.
And NVIDIA’s Open Agent Safety Platform provides a new example of how those principles can translate into a practical architecture where autonomous agents operate inside independently enforceable boundaries.
Together, these developments point toward a security model in which AI is both:
a capability for defending systems, and a system that must itself be continuously defended.
“As AI agents gain access to more tools, data and infrastructure, capability alone is not enough. They need independent controls around how they operate and evidence that the intended outcome actually occurred. Joining the Open Secure AI Alliance builds on our commitment to collective cyber defense and aligns with how we are developing Cortex: intelligence, execution, security and verification working as separate but coordinated layers.”
Anoop Jaishankar, Founder and CEO, Pervaziv AI
Building the Trust Layer for Agentic AI
The next phase of enterprise AI will depend on more than increasingly capable models.
Organizations will need to know what an agent can access, what it attempted to do, whether it remained inside policy, what changed during execution and whether the final result can be independently verified.
NVIDIA describes this as creating a trust layer for agents, drawing a parallel to the security technologies that allowed the internet to evolve from an open but risky environment into infrastructure capable of supporting global commerce and critical services. NVIDIA Developer
That is also where the opportunity for collective security becomes larger.
Through its membership in the Linux Foundation and Open Secure AI Alliance, alongside its support for OpenAI’s collective cyber defense initiative, Pervaziv AI is expanding its participation in efforts to build that trust layer across software and agentic AI.
As Cortex evolves across development, security, enterprise workflows, browsers and managed execution, Pervaziv AI will continue working toward the same objective:
Enable AI to take on more meaningful work while giving organizations the security, control and verifiable evidence required to trust what it does.
