Cisco opens bug-finding models built to beat bigger AI

Cisco has released two open-weight models built to find known vulnerabilities in existing codebases, betting that smaller, specialized systems can outperform far larger rivals on a narrow security task. The models — Antares-350M and Antares-1B — are the first in Cisco’s new Antares family of security small language models, and are available on Hugging Face to vetted users only.

“We’re making sure we’re gating that and appropriately granting access,” DJ Sampath, Cisco’s senior vice president and general manager of AI software and platform, told The Register.

Cisco said it is working with academic and nonprofit organizations, along with smaller and public organizations' security teams, to provide access. A larger 3-billion-parameter Antares model will not be publicly released, according to Amin Karbasi, Cisco VP and chief AI scientist.

Because the models are small enough to run locally, Cisco argues they are a better fit for organizations that cannot send proprietary code to cloud providers for analysis. Sampath told The Register that using Antares still requires access to the source code, limiting the risk of misuse to situations where an attacker has already compromised an endpoint or service.

Recommended reading

Top AI models caught cheating in up to 14.1% of tests

Antares benchmark claims against Google and OpenAI

Cisco says its models match or beat dozens of larger systems on a new benchmark designed to measure how efficiently models identify security flaws in codebases. According to the company:

• Antares-1B outperforms Google’s Gemini 3 Pro
• Antares-1B is comparable to Z.ai’s GLM-5.2
• The unreleased Antares-3B outperforms GLM-5.2 and OpenAI’s GPT-5.5

Karbasi told The Register that Antares can process 500 repositories in 15 minutes, while “frontier models” take five hours. He also said the cost is less than $1, versus $100 to $150 for larger models.

“Antares is inherently not a chatbot. It is an investigator. It is a search engine. It has to find a very specific thing that might be a needle in a haystack, and it goes and finds that.”

Karbasi said Cisco trained the models to try multiple search strategies when hunting for vulnerabilities, rather than relying on a single path. He compared that approach to a bicycle moving through a crowded London street faster than a truck.

The family’s name comes from Antares, the red supergiant star. Karbasi said the analogy is about scale: a vulnerability’s impact can be enormous even if it comes down to a single file or a few lines of code in a million-line codebase.

Sampath offered the simpler version of the pitch: sometimes, he said, “you don’t need a private jet to go to a corner store.”