Cisco launches accessible AI models that scan source code for security failures

Cisco revealed the family Antares, a set of small open-weight language models designed specifically to detect vulnerabilities in code bases. The initiative allows companies to perform the analysis locally, eliminating the need to send sensitive code to external AI providers — a growing concern for sectors such as financial, health and governments.

Antare models for code security: what are they?

The Antares models are compact versions of language models, trained to identify security failures in different programming languages. Because they are open weight, they can be downloaded, inspected and adapted by any development team. According to SecurityWeek, Cisco prioritized economy and privacy: models run on modest hardware (even conventional CPUs) and do not require cloud connection.

Antares vs. traditional code security solutions

Commercial code analysis tools usually charge per line or scan, and many send the code to remote servers. Antares, by running on-premises, eliminates transfer costs and leak risks. Although its reduced size may imply less precision in very specific cases, Cisco states that they achieve competitive detection rates for the most common vulnerabilities (SQL injection, buffer overflow, etc.). This makes them ideal for teams who want a first quick and cheap filter.

Why Brazilian developers should consider Antares

In Brazil, the General Data Protection Act (LGPD) imposes severe restrictions on sending personal data out of the country. Companies dealing with proprietary code often hesitate to use foreign SaaS tools. Antares offers an alternative that respects data sovereignty and reduces license costs. The trend of smaller models and open source also democratizes access to offensive security, allowing start-ups with lean budgets to integrate automated checks.

Impact of Antares on the AI market and code security

Cisco's move reinforces running for efficient models that can run on edge computing or on local servers. While giants like NVIDIA focus on high performance hardware for large models (such as NVIDIA Vera Rubin NVL72 reaches production: 10x more energy efficiency for scale AI), the bet on lean models shows that the industry recognizes the demand for affordable and controllable solutions. Code security has never been so critical – and having tools that do not compromise privacy is a competitive differential.

Frequently Asked Questions About Antares Models

Do Antare models replace tools like SonarQube or Fortify?

Not entirely. Antares is a complement that offers fast and economical scans, but does not replace advanced static and dynamic analyses. Ideal for teams that want an initial scan without relying on cloud.

Do I need special hardware to run the models?

Nope. The models were optimized to run on common CPUs, although GPUs accelerate inference. Cisco recommends at least 8 GB of RAM for the larger version, but there are even lighter variants.

Does the code really remain secure?

Yeah. As models run locally, no data is sent out of the corporate environment. In addition, open weight allows independent audit of what the model actually does.