Cisco Releases Antares: Open-Weight Models for Vulnerability Localization

Cisco released Antares-350M and Antares-1B, two open-weight small language models that locate where known vulnerabilities live inside a codebase. Both are free on Hugging Face and can run locally, allowing organizations to keep proprietary code inside their own environment. Antares finds candidate files. It does not generate fixes. A larger version, Antares-3B, is coming later.

What Cisco Announced

Cisco’s Foundation AI team released Antares on July 21 with the Vulnerability Localization Benchmark, a new benchmark built for this task. The models use a vulnerability description, linked to a Common Weakness Enumeration (CWE) category or advisory, to search a code repository like an analyst would. They review candidate files, follow naming conventions and call paths, change direction when evidence is weak, and narrow the search to the files most likely to contain the flaw. The result is a ranked file list and the search trace behind it.

Foundry Security Spec, a blueprint for building agentic security evaluation systems, and CodeGuard, a set of secure coding rules meant to guide AI coding agents toward safer output, were both released previously, CodeGuard a while back. Cisco released Antares as part of its broader open-source strategy to democratize AI access and tools for security; it is the model layer in that stack.

Small by Design

The compact size is the point. A 350 million or one billion parameter model costs a fraction of what a frontier-scale model charges to run the same localization task, and it is small enough to run on a laptop or an on-premises server. That matters for organizations that cannot send proprietary source code to an external API for compliance or contractual reasons and for teams that have never had the budget to run token-intensive AI models against every commit.

Cisco frames the natural users as universities, public sector institutions, and smaller security teams priced out of AI-assisted triage. A large enterprise with an established static analysis pipeline gets a different kind of value: a fast, always-on first pass that narrows what a human analyst looks at next.

Where This Fits in the Toolchain

Antares does one job. It does not replace software composition analysis, secret scanning, dynamic testing, infrastructure and container checks, or the expert review an analyst still must run before turning a finding into a fix. Cisco is explicit about this in its own release. The value is speed at the localization step, the part of triage that consumes analyst hours searching through unfamiliar code before any real remediation work starts.

Practical entry points include continuous integration and continuous delivery (CI/CD) pipelines that need an early view of which files are most likely to contain a relevant vulnerability and need early review, advisory-driven investigations after a public Common Vulnerabilities and Exposures (CVE) disclosure, and CWE-specific sweeps across large legacy codebases where teams lack a fast way to know where to start looking.

An Early Benchmark, Awaiting Independent Validation

Cisco built and published the benchmark behind its performance and cost claims – Antares outperforming larger closed and open models at a fraction of the cost – in the same announcement. Independent labs have not yet reproduced these results, so they remain a strong first claim rather than a settled one. The two outside endorsements, from professors at the National University of Singapore and Stanford, add credibility to the underlying research direction; they are not, however, independent verification of the specific numbers. Third-party benchmarking is the natural next step, and one worth watching for.

Two areas are worth watching as the model matures: false positive rates and performance against obfuscated or adversarially modified code – an increasingly realistic scenario as agents write more code. Security and compliance teams should also weigh a related concern: running any third-party model against production source code, even locally and free of charge, adds a vendor-risk and shadow-AI consideration alongside the data-residency benefits it provides.

Our Take

The open-weight release is more important than the benchmark result. Cisco is encouraging adoption of Foundry Security Spec and CodeGuard, the framework and rules layer where it wants to shape how agentic security tools are built and evaluated. Antares itself is not for sale. Offering a capable small model for free is a low-cost way to get developers working within Cisco’s specification instead of a competing one. The success of that strategy depends less on Antares’ model weights and more on whether the surrounding ecosystem gains broad adoption.

For security teams, the key question is whether Antares reduces the analyst time now spent finding vulnerable files, compared with the static analysis tools already in use. Neither Cisco nor an independent third party has run that comparison yet, so security teams evaluating Antares need to run their own version of it. Does your CI/CD pipeline have a localization gap that a free, locally run model could close faster than your current static analysis tools? If sending code to an external application programming interface (API) creates a compliance issue, Antares' local deployment model is worth piloting regardless of the benchmark debate. What would it take to test Antares against your existing SAST tool on your own codebase instead of relying on Cisco's published numbers?

Want to Know More?

Your Security Tools Were Built for People. Agents Are Not People.

Cisco Makes Its Bid to Run the Agentic Enterprise at Cisco Live US 2026

Why Cisco, Why Now: How Cisco’s RSAC 2026 Strategy Reframes Agentic AI Risk for CISOs

The 2026 Turning Point: Key C-Suite Takeaways From the Cisco AI Summit