Cisco has released Antares, a new set of security-focused small language models (SLMs) aimed at helping locate known vulnerabilities within software codebases.
The announcement was made by Cisco Foundation AI vice president Amin Karbasi, in a company blog. Two Antares models, Antares-350M and Antares-1B, are now available as open-weight models on the Hugging Face platform.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
The Antares models are designed for vulnerability localisation in source code, described by Karbasi as “one of the hardest, most time-consuming and expensive problems in security.”
Cisco claims its internal benchmarks show that these compact models provide better performance in code vulnerability identification than several existing closed and open-weight models, and do so at lower cost. Both models can be run locally, allowing organisations to analyse sensitive code on-premises rather than in the cloud.
Antares forms part of a wider Cisco strategy to make AI-supported security tools more accessible and practical.
Karbasi wrote: “By combining open-weight models with Antares, open specifications with Foundry Security Spec, secure coding guidance with CodeGuard, and a new benchmark, we are working to define the types of practical, trustworthy AI tools that will help cybersecurity professionals do their job better, as well as cultivating an ecosystem and standards for further enterprise adoption of AI security tools.”
According to Cisco, effective code security depends on connecting external vulnerability knowledge, such as public databases, advisories, and Common Weakness Enumeration (CWE) data, to internal source code. This connection is complex due to large codebases and dispersed evidence.
Cisco has engineered Antares to support security professionals in reviewing unfamiliar code, handling complex naming conventions, and tracing potential weaknesses.
The company positions Antares as an option for organisations seeking practical security tooling without extensive resources. The smaller size of the models is intended to reduce computing costs and allow for local deployment, enabling organisations, including universities and public sector bodies, to keep sensitive source code in-house while using AI assistants.
The Antares models are designed to improve on traditional static analysis tools, which often generate large volumes of results requiring analyst triage. Cisco maintains that existing general-purpose code models lack features for structured vulnerability search or repository navigation.
Antares aims to bridge this gap by applying compact models that adopt an iterative, investigator-like search pattern. The process starts from a vulnerability description, searches for candidate code, reviews files, and revises its approach as it gathers evidence.
Karbasi said that Antares does not replace expert review but is intended to make the early steps of triage more efficient and consistent. The models produce a ranked list of source files suspected of containing vulnerabilities, together with detail on how those files were identified.
The output can be used to prioritise review during advisory triage, CWE investigations, or code checks in development pipelines. Cisco sees Antares as complementary to broader security processes such as dependency analysis or dynamic testing.
Along with Antares, Cisco released the Vulnerability Localization Benchmark, a 500-task dataset to assess how well models can find vulnerabilities in unfamiliar codebases. Cisco claims that existing industry benchmarks do not address security-specific localisation tasks in the same way.
The release of Antares as open-weight models reflects the aim of lowering barriers for organisations with limited resources to adopt practical, AI-driven vulnerability detection, Karbasi wrote.
The Antares launch is aligned with other Cisco initiatives including Foundry Security Spec and CodeGuard, which focus on building accessible and reviewable AI security tools for enterprise use.
Last month, Cisco introduced Cisco Cloud Control, a platform designed for both human users and AI agents to manage, monitor, and secure critical IT infrastructure. Cisco said Cloud Control provides a unified interface for networking, security, compute, observability, and collaboration tools, giving users and AI agents a consolidated operational view.
