Cisco Launches Antares AI Models for Vulnerability Detection in Large Codebases

Cisco has announced a new family of small language models (SLMs) called Antares. These new AI models are designed specifically to help security teams identify known vulnerabilities within large codebases.

According to Cisco, security teams often find it challenging to connect a publicly reported software vulnerability to extract files that contain the flaw inside a large and often unfamiliar codebase. They must manually examine thousands of files, follow complex code relationships, and interpret scattered clues before they can determine whether a vulnerability actually exists in their software. This process is slow, expensive, and requires highly specialized expertise.

Antares models deliver enterprise AI without cloud dependency

Cisco has launched the first models called Antares-350M and Antares-1B, with a larger Antares-3B model planned for later. The Antares range focuses on identifying which source files are most likely to contain a known vulnerability based on security advisories, CWE descriptions, and other vulnerability information.

“Benchmark testing shows that these models outperform many powerful closed- and open-weight models in this critical security task at a fraction of the cost,” explained Amin Karbasi, VP and Chief Scientist for Cisco Foundation AI. “And they’re compact enough to run locally, heading off the need to send sensitive codebases to the cloud.”

Cisco highlighted that these small language models can match or outperform many larger open- and closed-weight AI systems on vulnerability-localization tasks while requiring fewer computing resources and lower operating costs. Moreover, enterprises can run these AI models on-premises instead of sending proprietary source code to cloud services. This approach helps to improve privacy and makes the technology more accessible to organizations with limited budgets or strict compliance requirements.

Antares follows an iterative search approach rather than relying entirely on fixed rules. It starts from a vulnerability description, explores the repository, examines candidate files, gathers evidence, revises its search path when needed, and gradually narrows the investigation to the most relevant files. The models produce a ranked list of potentially vulnerable files as well as the search trail used to reach those conclusions.

“We are releasing Antares as open-weight because the security community needs more accessible building blocks for practical, repository-level defense,” said Karbasi. “Cisco’s efforts are connected by a common belief: AI in security has to move beyond impressive one-off demos and toward systems that practitioners can evaluate, govern, and improve.”

Cisco VLoc Bench measures AI vulnerability detection accuracy

Cisco has also introduced the Vulnerability Localization Benchmark (VLoc Bench). It’s a 500-task evaluation framework created to measure how effectively AI systems can detect vulnerabilities in unfamiliar codebases. Cisco claims that existing coding benchmarks don’t properly test this specialized security task.

The launch of Antares is a part of Cisco’s broader effort to develop practical and trustworthy AI tools for cybersecurity professionals. The company has also introduced other initiatives such as security specifications, secure coding guidance, and open evaluation standards.