Key Takeaways
- The Open Secure AI Alliance now includes over 120 organizations collaborating on agentic AI cybersecurity.
- A new Request for Comments, the Shared AI Findings Exchange (SAFE), proposes confidential collection and analysis of AI incidents to generate shared, evidence‑based defenses.
- NVIDIA contributes a full‑stack of open tools—from the NOOA research harness and OpenShell runtime to verified agent skills, safety guardrails, and the Garak vulnerability scanner.
- Alliance members such as Okta, Palo Alto Networks, Red Hat, Amazon, Capital One, Cloudflare, Microsoft, Cisco, CrowdStrike, Akamai, Cognition, Perplexity, Uber, LangChain, and Veeam are adding identity, harness, model, observability, and resilience capabilities across the AI security stack.
- The alliance invites participants to a Black Hat group photo session on Aug 4 at 5:15 pm PT and encourages further involvement in open AI security efforts.
Introduction and Alliance Overview
Members of the Open Secure AI Alliance—now exceeding 120 organizations—are intensifying their work to harden agentic AI cybersecurity as the annual Black Hat conference kicks off in Las Vegas. The Linux Foundation has released a Request for Comments on the Shared AI Findings Exchange (SAFE), a proposed guideline set designed to turn agentic AI security incidents into collective protection for the entire ecosystem. The SAFE draft is being prepared by an Open Secure AI Alliance working group, with active contributions from NVIDIA, Cisco, CrowdStrike, Hugging Face, and Red Hat, among others. By pooling threat intelligence and lessons learned, the alliance aims to move defenses at “agent speed” to keep pace with the rapidly evolving attack surface created by AI technologies.
SAFE Guidelines Overview
The SAFE framework outlines a process for confidentially gathering and analyzing AI incidents and near‑misses, informing affected parties, identifying recurring control failures, and publishing evidence‑based operating recommendations that lower systemic risk. Recognizing that cybersecurity is a race without a finish line, the alliance argues that open, timely sharing of threat intelligence transforms individual defenses into a force multiplier. When organizations disclose what they observe—whether a successful breach, a thwarted attempt, or a anomalous behavior—the community can derive patterns, improve mitigations, and raise the baseline security posture for all participants in the AI supply chain.
NVIDIA Full‑Stack Contributions
NVIDIA’s contributions span the entire AI security stack. The NVIDIA Labs Object‑Oriented Agent (NOOA) research harness, hosted on GitHub, simplifies testing, tracing, auditing, and governing agent behavior. Complementing this, the NVIDIA OpenShell runtime enforces security and privacy controls at the agent level, restricting what an agent can see, touch, or do. NVIDIA also releases open model families—Nemotron for agentic AI, Cosmos for physical AI, Isaac GR00T for robotics, BioNeMo for healthcare, and Alpamayo for autonomous vehicles—each accompanied by open weights, datasets, and training procedures. At the capability layer, NVIDIA provides verified agent skills: portable instruction sets that are cryptographically signed, scanned for prompt injection and tool poisoning, and documented with skill cards so defenders know exactly what a skill does, its provenance, and any post‑publication modifications. Additional tools such as NeMo Guardrails, NeMo Anonymizer, NeMo Safe Synthesizer, and the open‑source Garak LLM vulnerability scanner help enforce safety policies, protect sensitive data, generate privacy‑safe synthetic data, and scan models for data leaks, jailbreaks, and prompt injections before deployment.
Identity and Permissions Contributions
Securing agents begins with knowing who—or what—is acting. Okta is developing reference implementations for agent identity and access, demonstrating how the Cross App Access (XAA) open protocol lets AI agents running in OpenShell sandboxes safely connect to enterprise applications. Palo Alto Networks contributed open‑source tools from its Idira identity platform, including Agent Guard and Agent Watch, which help developers apply identity‑security best practices and safely retrieve secrets for agentic workflows. Red Hat’s new asago project maps an organization’s custom governance requirements—drawn from NIST, OWASP, and the EU AI Act—to runtime agent permissions, producing a single audit trail that links policy clauses to live controls. Together, these efforts aim to provide fine‑grained, verifiable boundaries that prevent unauthorized actions while preserving legitimate agent functionality.
Harnesses and Tooling Contributions
If a model is the agent’s brain, the harness is its body, orchestrating tool use and governing behavior. Amazon, a recent alliance member, offers Strands Agents—an open‑source toolkit for building AI agents that is transparent at every layer, enabling full visibility into agent behavior and production‑time evaluation. Amazon also supplies Cedar, an open‑source authorization language that enforces deterministic, verifiable limits on what agents may do, delivering fine‑grained, analyzable access controls. Capital One open‑sourced VulnHunter for agentic AI code security, while Cloudflare provides its Vulnerability Discovery Harness as an open‑source skill to harden agent systems. Microsoft’s AI Red Team contributed PyRIT (Python Risk Identification toolkit) for automated red‑teaming, RAMPART to turn red‑team findings into repeatable tests, Clarity to question design assumptions, and Assert to convert natural‑language safety requirements into executable evaluations. Wiz’s Atlas autonomously researches vulnerabilities using specialized AI agents, and Visa contributed its open‑sourced Visa Vulnerability Agentic Harness to streamline issue identification, remediation, and validation.
Specialized Security Models
Not every safety task needs a general‑purpose model; purpose‑built security models increase efficiency and precision. Cisco released DefenseClaw, an open‑source agentic governance layer that sits atop NVIDIA OpenShell to deliver automated runtime security at scale, plus two Antares small language models that pinpoint known vulnerabilities in codebases and Project CodeGuard, which embeds secure‑by‑default practices directly into AI coding workflows. CrowdStrike is fine‑tuning the NVIDIA Nemotron Nano model for cyber defense, achieving 96 % accuracy in generating investigation queries within Falcon LogScale and demonstrating that a specialized Nemotron Nano reasoning model outperforms larger models in SOC detection triage while providing calibrated logit‑based confidence for auditable autonomous decisions. Akamai feeds insights from its State of the Internet reports and Security Intelligence Group research to illuminate AI‑era threats, while Cognition’s trustworthiness evaluation measures alignment and security risks of open‑source‑derived models, showing how post‑training can mitigate those risks. Perplexity’s Numbat offers an open‑source agent security suite for macOS, Linux, and Windows endpoints, detecting, investigating, and preventing agent activity and delivering a structured record of agent behavior. Uber’s open‑sourced ADR (Agentic AI Detection and Response) reconstructs the full causal chain of agent activity—from prompt to reasoning, tool calls, and outcomes—supporting over 200,000 agent sessions per day across 30,000 endpoints via a two‑tier detection‑and‑investigation approach.
Observability, Evaluation, and Threat Intelligence
Effective defense requires visibility into what agents do and why they act. Beyond raw logs, defenders need explanations of intent, safety assessments, and awareness of evolving attack techniques. The alliance’s contributions in this area include Akamai’s threat‑intel feeds, Cognition’s trustworthiness scoring, Numbat’s endpoint‑level monitoring, and Uber’s ADR causal‑chain reconstruction. These tools give security teams a multi‑dimensional view: they can see actions, understand motivations, verify compliance with safety policies, and detect subtle shifts in adversary tactics. By integrating observability with automated evaluation (e.g., Microsoft’s Assert and RAMPART), organizations can continuously validate that agents adhere to prescribed security postures and rapidly adapt when deviations are detected.
Availability and Resilience Enhancements
Agent systems must stay dependable even when disruptions occur, containing failures and recovering safely without losing critical state or exposing the broader environment. LangChain is adding resilience capabilities to its open‑source frameworks—Deep Agents, LangGraph, and LangChain—allowing agents to retry interrupted work, follow safe recovery paths, resume from saved states instead of restarting, and automatically fall back to alternative models when the primary model fails. Veeam contributes Kanister, an open‑source framework for data protection on Kubernetes, helping teams safeguard and recover AI workloads, vector databases, and associated data to a verified known‑good state. These mechanisms ensure that AI services remain available and that any incident can be contained and remediated with minimal collateral impact.
Call to Action at Black Hat
The Open Secure AI Alliance invites all interested parties to join its Black Hat gathering today, Tuesday, Aug 4, at 5:15 pm PT, for a group photo outside the Main Stage, Business Hall at the Mandalay Bay Convention Center. This event offers a chance to meet fellow contributors, learn about the latest open tools and guidelines, and explore how to participate in the alliance’s ongoing work. Those wishing to learn more or express interest in joining can visit the alliance’s website or contact the Linux Foundation for details on membership, contribution pathways, and upcoming collaborative sessions.
Conclusion and Invitation
As AI agents become integral to enterprise operations, securing them demands a holistic, community‑driven approach that spans identity, harnesses, models, observability, and resilience. The Open Secure AI Alliance’s expanding catalog of open tools—highlighted by NVIDIA’s full‑stack offerings and the diverse contributions of partners such as Okta, Amazon, Microsoft, Cisco, CrowdStrike, and many others—demonstrates that collective defense is not only possible but already underway. By adopting frameworks like SAFE and leveraging the shared mitigations produced by the alliance, organizations can turn individual security gains into ecosystem‑wide protection, ensuring that the promise of agentic AI is realized safely and responsibly.