Cisco releases Antares AI models to find vulnerabilities

Cisco has released a pair of specialized artificial intelligence models designed to help security teams quickly locate potential software vulnerabilities within large codebases. The open-weight models, named Antares350M and Antares1B, were made available on the Hugging Face platform on July 21, 2026. The company is positioning them as a cost-effective and private alternative to large, general-purpose AI systems for the initial triage of security flaws.

The Antares models are built for a specific task known as vulnerability localization. When a new security weakness is publicly disclosed, analysts must determine if their organization's proprietary code contains a similar flaw and pinpoint its exact location. This process is often slow and labor-intensive. Antares aims to accelerate it by acting as a search agent. Given a description of a vulnerability, such as a Common Weakness Enumeration (CWE) entry, the model systematically explores a code repository, reads files, gathers evidence, and outputs a ranked list of the source files most likely to contain the issue. This allows human security engineers to focus their investigation on a narrowed set of candidates rather than...