Cisco releases open-weight Antares models for finding vulnerable code files

Cisco has introduced Antares, a new family of small-language AI models aimed at a narrow but costly security problem: finding the files in a large codebase that are most likely to contain a vulnerability. The company says the first two Antares models are now available as open-weight releases, with the work positioned as a lower-cost, locally deployable alternative to sending sensitive source code to a frontier model in the cloud.

The announcement, published by Cisco on August 4, frames Antares as a tool for analysts who are trying to triage large repositories where the useful signal may be scattered across many files. Rather than replacing broader AI coding assistants or remediation tools, Cisco says the models are trained to explore repositories, gather evidence, revise their search, and return a ranked list of files for security teams to inspect.

Why it matters

Software teams are leaning harder on code-generating systems, but the security review burden has not disappeared. Cisco argues that smaller, specialized models can help organizations that lack the budget or infrastructure for repeated frontier-model analysis, especially universities, nonprofits, public-sector groups, and smaller teams with large codebases. Local deployment is also central to the pitch: source repositories can reveal product architecture, intellectual property, security controls, and customer or operational details.

Cisco described three Antares model sizes. It said the 1B model is the primary practical release for laptops and workstations, while a larger 3B model remains gated because the company views it as powerful enough to require controlled access. The first two releases are being made accessible through Hugging Face, according to Cisco's newsroom article.

The bigger trend

Antares fits a broader shift in security AI from all-purpose assistants toward narrowly scoped models that can run near the data they analyze. That does not remove the need for expert review; a ranked file list is still a starting point, not a confirmed vulnerability report. But if the models perform as Cisco claims, they could make vulnerability localization more routine for teams that cannot repeatedly rely on expensive hosted systems or broad manual audits.

The release also underscores a tension in open AI security tooling. Wider access can help defenders test and improve software, while more capable vulnerability-finding systems can create misuse concerns. Cisco's decision to open some weights while gating the larger model reflects that balance.

Sources

• newsroom.cisco.com

Cover photo by Markus Spiske on Pexels, used under the Pexels License.