Open-Source Coalition Rewires AI Security for the Agentic Era

Databricks Joins an Open-Source Coalition to Rewire AI Security from the Harness Upward

Databricks has joined the Open Secure AI Alliance as a founding member, contributing a suite of open-source frameworks and tooling to advance AI safety and security through open research.

The move places Databricks alongside Nvidia, CrowdStrike, and more than 75 other organizations in a coalition that pledges to build and share open models, harnesses, and threat intelligence for securing both AI systems and AI-enabled cyber defenses.

The alliance’s approach addresses a fundamental gap: as AI agents grow autonomous, the execution stack — the harnesses, guardrails, and governance frameworks — must be open to inspection if enterprises are to trust autonomous decisions.

Omnigent, BlackIce, and the Frameworks That Fortify Agentic AI

While much of the AI safety discourse has centered on model weights, the Open Secure AI Alliance targets the full agent stack — from the runtime harness that orchestrates model and tool interactions to the governance layer that audits every decision.

As Databricks puts it,

The alliance is built on a simple but powerful premise: AI safety and security research should be shared openly, and the tools it produces should be built on open systems.

Among the inaugural technical contributions, four open-source pillars stand out.

• Omnigent: An open-source meta-harness under Apache 2.0 that composes and securely shares agents across 13 harnesses while enforcing spend caps, contextual policies, and sandbox isolation. It integrates with NVIDIA OpenShell for kernel-level isolation.
• DASF 3.0: The Databricks AI Security Framework catalogues 97 technical risks across 13 components and maps them to 73 mitigation controls, aligned to MITRE ATLAS, NIST, and other standards. Version 3.0 adds dedicated coverage for agentic threats like memory poisoning and insecure MCP connections.
• DAGF: The Databricks AI Governance Framework defines accountability structures across five pillars, helping enterprises govern AI programmatically and align open technology adoption with compliance requirements.
• BlackIce: An open-source red-teaming toolkit that bundles 14 security tools into a reproducible container environment, already downloaded over 6,000 times. It lowers the barrier for teams to test for prompt injection, data leakage, and supply-chain attacks.

Lakewatch, an open security lakehouse architecture, provides an open, governed data layer for agentic threat detection and response at scale.

Unity Catalog, open-sourced under Apache 2.0 and donated to the Linux Foundation, governs which agent accesses which data, model, or tool — keeping the most critical control point auditable and portable.

As Nvidia’s Open Alliance Gains Steam, Absent Players Signal a Fractured Safety Landscape

The initiative’s roster, now exceeding 75 members including Microsoft, Red Hat, and Palantir, signals broad industry appetite for open security frameworks. Yet the omission of OpenAI, Anthropic, and Google — all absent from the founding list — reveals a deeper ideological split.

The Hill reported that Nvidia launched the alliance, emphasizing that open models ‘democratize defensive capabilities’ and guard against single points of failure. Nvidia’s push coincides with a joint letter from Microsoft, Meta, and others urging policymakers not to overregulate open-weight AI models.

CrowdStrike, an inaugural partner, shared data that crystallizes the practical argument. In vulnerability research, a generic harness produced false positives near 80 percent. A security harness developed by CrowdStrike dropped that to approximately 20 percent, while still surfacing more than 2,400 vulnerability findings across 43 products in days.

CrowdStrike also fine-tuned an open Llama Nemotron Super 49B model for natural-language-to-CQL translation, achieving 96 percent valid query accuracy — outperforming tested closed-source frontier models on that specialized task.

Entrenched resistance from frontier labs is not just philosophical. The Hill noted that Anthropic, along with CEO Dario Amodei, has consistently opposed open-weight models, citing existential misuse risks.

Yet Nvidia’s counterpoint is sharp: closed systems do not eliminate misuse, as shown when OpenAI models breached Hugging Face’s systems in a test environment. Hugging Face deployed the open-weight GLM 5.2 model on its own infrastructure to analyze 17,000 actions and neutralize the intrusion — a sequence that would have been impossible if all components were locked behind proprietary walls.

Why the Agentic Era Demands Transparency from Silicon to SOC

Security in the age of autonomous agents can no longer be a bolt-on checkbox; it must be woven into the fabric of the execution stack — from the harness that controls tool use to the catalog that governs data access. The Open Secure AI Alliance makes a bet that opening every one of those layers to inspection will accelerate defensive innovation far faster than proprietary secrecy.

The alliance’s early signals — tangible frameworks, reproducible benchmarks, and a growing community of contributors — suggest that enterprises can now build and certify AI security on an open foundation, rather than leasing it from a handful of vendors.

For organizations navigating the shift to open, composable architectures, the same principle of inspectable, high-performing systems applies to their digital presence. Andres SEO Expert engineers AI-powered programmatic SEO strategies that turn open-source capabilities into measurable traffic growth. To discuss a tailored roadmap, connect with Andres and learn more about Andres SEO Expert’s approach to building resilient, automated marketing engines.