Cisco released Antares this week — a family of compact, open-weight AI models designed for one job only: finding security vulnerabilities in code. And the numbers are hard to ignore.
The first two models, Antares-350M and Antares-1B, are available now on Hugging Face as open-weight releases. A larger Antares-3B model is coming soon. Cisco says a full 500-entry evaluation with Antares runs in about 15 minutes on a single GPU and costs under $1. That’s 15 times cheaper than the best open model they tested against, and 172 times cheaper than the leading frontier model in the same benchmark — and for context, Microsoft recently used AI to patch 622 flaws in a single day.
For someone like me who spends a good chunk of each week thinking about both AI and security — and the uncomfortable gap between them — this feels like a genuinely important release.
Not Another General-Purpose Coding Assistant
Here’s what makes Antares different from the dozens of coding-focused models hitting Hugging Face every week: it doesn’t try to do everything.
Most AI coding tools aim to write code, explain code, refactor code, and debug code all at once. They’re generalists. Antares is a specialist. It takes a vulnerability description — say, a CWE identifier or a security advisory — and searches through a codebase to find the files most likely to contain the relevant weakness. That’s it. It outputs a ranked list of source files and a trace of how it got there.
This kind of focused localization is surprisingly hard. Static analysis tools throw too many false positives. Frontier models are powerful but expensive and require sending proprietary code to the cloud. A compact model that runs locally, costs pennies per scan, and narrows down thousands of files to a handful of candidates is exactly what most security teams actually need.
The Smartest Thing About Antares? The Benchmark
Cisco didn’t just drop models and walk away. They also built the Vulnerability Localization Benchmark (VLB) — a 500-task dataset that measures how well models can navigate unfamiliar codebases and identify files associated with specific CWE categories.
This matters because most coding benchmarks test general software engineering skills: can the model fix a bug, write a test, or search for a function. Security localization is a different skill entirely. You’re asking the model to understand what a “CWE-79 Cross-Site Scripting” vulnerability looks like in practice, then hunt through a real repository and point to the files where it might live.
On that benchmark, Antares-3B (coming soon) achieved a File F1 score of 0.223 — close to GPT-5.5 at 0.229. The Antares-1B model posted the highest recall among all evaluated systems. And the tiny Antares-350M outperformed several much larger general-purpose models. That’s remarkable for models small enough to run on a laptop.
Why Open-Weight Matters for Security
Cisco is releasing Antares as open-weight, not just as an API. That distinction matters more in security than in almost any other domain.
When you’re analyzing proprietary source code for vulnerabilities, sending that code to a third-party API creates real risk. Even with data processing agreements and encryption, the act of transmitting your codebase to an external service is a compliance and trust boundary that many organizations — especially in government, finance, and critical infrastructure — simply won’t cross. How to audit AI coding tools for data privacy covers the practical steps for keeping your code safe.
Antares runs entirely on-premises. Air-gapped environments, closed networks, and compliance-heavy deployments are all viable use cases. For a government ICT division like the one I work in, that’s the difference between a tool we can actually use and one we have to admire from a distance.
Cisco’s Bigger Security AI Bet
Antares isn’t an isolated product drop. It’s part of a broader push from Cisco that includes two other important pieces:
Foundry Security Spec — a model-agnostic framework for building agentic security evaluation systems with clear roles, guardrails, and reviewable outputs. It’s Cisco saying that the way AI interacts with security workflows needs to be governed, not just deployed.
CodeGuard — an open-source framework that embeds secure-by-default practices into AI coding agent workflows. Think of it as a rules engine that tells your AI coding assistant “before you generate that function, check if it follows these security patterns.”
Put together, AI-generated code vulnerabilities are a growing concern. Cisco is building a stack: CodeGuard prevents vulnerabilities at generation time, Antares catches what slips through, and Foundry Security Spec makes the whole thing auditable. It’s a surprisingly coherent vision from a company most of us still associate with routers and switches.
What This Means for Developers and Security Teams
The realistic takeaway here isn’t that Antares will replace human analysts or static analysis tools. It won’t, and Cisco doesn’t claim it will. What it does is compress the first stage of vulnerability triage — the part where you stare at a CVE advisory and wonder which of your 500 microservices might be affected — from hours to minutes.
For smaller teams, the economics matter most. Security shouldn’t be a luxury that only well-funded enterprises can afford. A model that costs under $1 to evaluate 500 files, runs on a single GPU, and doesn’t require shipping your code to someone else’s server changes the conversation about what’s practical.
For larger organizations, the appeal is different. Running compact models on every commit — not just before releases — becomes feasible when each scan costs pocket change. That shift from periodic security reviews to continuous, commit-level scanning is where AI-assisted security really starts to pay off.
Bottom Line
Cisco Antares isn’t going to single-handedly solve the software vulnerability problem. No model will. But it’s a rare example of an AI release that’s targeted, practical, and honest about what it does.
It doesn’t claim to be general intelligence. It claims to find vulnerable code, cheaply and locally. And based on the benchmarks, it does that better than models a hundred times its size. That’s not hype. That’s actually useful.
The Hugging Face page for Antares is live now. If you’re responsible for security in your organization — or even if you just care about writing safer code — it’s worth a look.