Open Secure AI Alliance Expands at Black Hat: What You Should Know

The Open Secure AI Alliance expands at Black Hat with SAFE guidelines and open frameworks, offering IT leaders inspectable tools, lower token costs, and a clear roadmap for secure enterprise AI.

Technologies scale much faster when they are open and interoperable. This has been true for networking, VoIP, mobile, and almost every other technology shift over the past few decades. The Open Secure AI Alliance is a significant step toward making AI defense open, inspectable, and enterprise-ready, and it arrives at a moment when IT teams are trying to turn pilot projects into production value.

Why an open AI defense stack matters

Open source has long been the backbone of modern infrastructure, spanning cloud and compute platforms, and cybersecurity is among its biggest beneficiaries. The Open Secure AI Alliance (OSAA) applies that same model to AI security by committing to shared open models, agent harnesses, tools, and data that defenders can study, extend, and run on their own infrastructure.

The core argument behind OSAA is that if the next generation of cyberattacks is AI-powered, the next generation of cyber defense must be too, meaning defenders need systems they can actually see into. In a recent incident on Hugging Face, closed AI tools were blocked, preventing essential forensic analysis and forcing the team to rely on an open-weight GLM 5.2 model running on its own infrastructure to analyze more than 17,000 actions and contain the intrusion.

That episode sheds light on a new truth for CISOs and security operations leaders: if you can’t inspect and adapt your AI defenses in the moment of a breach, you’re blind.

Who’s in the alliance and why it’s different

OSAA brings together a broad roster spanning cloud, cybersecurity, enterprise software, and open-source foundations, among others. For enterprises, this matters because it signals that the AI defense stack won’t be a single-vendor vertical product but an ecosystem anchored in open standards and components.

Several of the founding contributions highlight how deep this stack can go:

• HPE is contributing to SPIFFE/SPIRE, advancing zero-trust identity standards that can cryptographically verify AI agents and services before they interact with enterprise resources.
• Hugging Face has introduced Safetensors, a safer model-weight format that prevents remote code execution and increases transparency in how models are stored and loaded.
• IBM and Red Hat’s Lightwell extends security across the open-source supply chain with digitally signed patches, helping close gaps in the software layer that AI systems rely on.
• Microsoft’s MDASH is a multi-model, agentic scanning harness that uses specialized agents to discover, debate, and prove the existence of exploitable bugs, rather than relying on a single opaque scanner.

These components form an emerging “open defense stack” for AI agents that spans identity, isolation, safe model formats, multi-model scanning, and secure development workflows. For IT leaders, this matters because it promises consistent controls across a multi-vendor environment, rather than brittle point tools that don’t interoperate.

Nvidia, Cisco and the role of agent harnesses

Nvidia is backing the alliance not only with models and weights but also with open agent-harness research through its Nvidia Labs Object-Oriented Agent (NOOA) framework. NOOA starts with a key insight for practitioners: model choice is only half the equation; the harness, the architecture around the model, can account for double-digit swings in benchmark performance and large differences in token cost, even when the underlying model remains unchanged.

In NOOA, an agent is represented as a single Python class. Its methods define capabilities, fields define state, docstrings serve as prompts, and type annotations enforce contracts. That design makes agent development look and feel like traditional software engineering: agents can be diffed, code-reviewed, unit-tested, and traced with the same tools teams already use for other critical systems.

For enterprises, this is a subtle but important shift as it turns agentic AI from a “black box workflow graph” into a testable, auditable code surface that security and platform teams can govern.

NOOA identifies six harness capabilities that materially affect model performance:

• Typed input and output, so calls have validated arguments and returns instead of free text.
• Pass-by-reference semantics, where the model operates on live Python objects and bounded previews instead of serialized dumps pushed into the context window.
• Code-as-action, meaning the model writes Python that invokes methods and control flow directly.
• Programmable loop engineering, with orchestration loops expressed as ordinary Python that humans and models can modify.
• Explicit object state, keeping durable typed state on the agent object instead of only in conversation history.
• Model-callable harness APIs for context blocks and event history, letting the agent inspect and manage its own working set.

For IT pros, harness engineering is important because it improves efficiency. In SWE-bench, NOOA achieves parity or better accuracy with roughly half the tokens of comparison harnesses, thanks largely to pass-by-reference and the ability to avoid costly context-compaction pipelines. For enterprises wrestling with LLM cost curves, that combination of higher accuracy and lower consumption directly accelerates time to value.

Cisco’s presence in the alliance builds on its broader partnership with Nvidia across networking, data center AI fabrics, and secure AI factories. Cisco has already committed to integrating technologies such as Silicon One and specialized SuperNICs into AI networking platforms and is advancing secure AI infrastructure from core to edge through offerings such as Cisco Secure AI Factory with Nvidia.

As these networking and security layers intersect with open harnesses and models from OSAA, IT teams gain a clearer path to end-to-end architectures in which compute, data, agents, and network security align with common, open principles.

Black Hat Update: The SAFE Guidelines and Expanding the Open Stack

As the annual Black Hat USA conference kicks off, the Open Secure AI Alliance announced a major new ecosystem initiative.

The Linux Foundation, in collaboration with alliance members, released a Request for Comments (RFC) on the Shared AI Findings Exchange (SAFE) guidelines.

SAFE is designed to turn agentic cybersecurity incidents and “near misses” into shared protection across the enterprise. Rather than keeping breach details locked behind non-disclosure agreements or confined to single-vendor ecosystems, SAFE provides a framework to:

• Confidentially collect and analyze AI security incidents, agent misbehaviors, and near-miss operational events.
• Inform impacted parties quickly while identifying recurring control failures in real time.
• Publish evidence-based operating recommendations to lower systemic risk for all defenders.

Alongside the SAFE framework, OSAA members are expanding the open software layer. Nvidia, for instance, introduced Nvidia OpenShell, an open runtime that serves as an agent-level sandbox. OpenShell limits what an agent can see, touch, or execute, enforcing explicit security and privacy boundaries so autonomous tools cannot access systems or data beyond their designated scope.

Together, these Black Hat announcements reinforce a key principle for enterprise security teams: securing agentic AI requires collective threat intelligence and open runtime guardrails, not just isolated model scanners.

Accelerating enterprise time to value for AI

For enterprise IT, the question is less “does this alliance matter?” and more “how does it change my roadmap?” There are several direct impacts on time to value:

• Faster, safer experimentation: Open models and harnesses that are inspectable and self-hostable let teams run high-risk security workloads, such as incident response or vulnerability discovery, on their own infrastructure without waiting for closed providers to support niche use cases.
• Reduced integration friction: An open defense stack spanning identity, format safety, scanning, and secure development provides a reference architecture for AI security, shortening design cycles and avoiding bespoke integrations for each tool.
• Better ROI from existing models: Techniques such as typed I/O, pass-by-reference, and structured memory can improve accuracy and reduce token usage with the same models, turning stalled pilots into production deployments that meet both performance and budget constraints.
• Stronger systemic resilience: Open, multi-vendor ecosystems reduce the risk of single points of failure or policy shocks that arise when a small set of closed providers controls critical defensive capacity.

In practice, this means AI security projects can shift from “hero-based” experiments to repeatable patterns. Teams can apply NOOA-style harness designs to their incident response agents, code reviewers, or threat-hunting workflows, and integrate those agents with identity, logging, and scanning layers aligned with OSAA contributions.

Guidance for IT and security leaders

For IT pros planning or scaling AI initiatives, several concrete steps can help them take advantage of this:

• Treat the alliance outputs as a reference architecture. As OSAA components mature, map them to your existing stack: SPIFFE/SPIRE for agent identity, Safetensors for model storage, Lightwell for supply-chain integrity, MDASH-like approaches for multi-model scanning, and harness patterns like NOOA for your agent platforms.
• Make harness engineering a critical discipline. Don’t frame agent projects solely as “model selection”; invest in harness design. That is, use typed contracts, pass-by-reference, programmable loops, and explicit states, and measure their impact on accuracy, latency, and token costs.
• Align security and development practices around agents as code. With NOOA, agents are Python classes that can be code-reviewed, unit-tested, and traced like any other critical service. Extend CI/CD, secure coding, and observability practices to these harnesses rather than treating them as one-off automation scripts.
• Advocate for open, inspectable defensive tools in policy and procurement. When evaluating AI security products or drafting internal policies, emphasize support for open models and harnesses that you can run and inspect on-prem or in your trusted cloud environments.
• Start with high-leverage security use cases. Early projects might focus on vulnerability discovery in your own codebase, automated configuration analysis, incident triage, or threat-hunting agents explicitly designed to work within NOOA-style harnesses and OSAA-aligned identity and logging.

The Open Secure AI Alliance’s core thesis is that AI defense need not be opaque or centralized. With the right combination of open models, rigorous harness engineering and shared defensive tooling, enterprises can build AI systems that are not only powerful but also observable, testable and under their control.

For IT leaders under pressure to deliver AI outcomes without compromising security, that is a future worth embracing now.

Other News: Google is tightening Chrome extension policies to curb abuse by restricting high-risk permissions, strengthening enterprise controls, and making it harder for malicious add-ons to exploit browser access.