Cisco Antares: Open-Weight AI That Scans Your Code for Security Flaws, Locally

Cisco Antares: Open-Weight AI That Scans Your Code for Security Flaws, Locally

The direct answer: Cisco has released an open-weight AI model family called Antares in three small sizes — 350 million, 1 billion, and 3 billion parameters — built for one specific job: identifying which files in a real code repository contain a security vulnerability. The license is Apache 2.0, so commercial use is permitted, and the models are small enough that the largest runs on a single GPU. The practical result for a business owner: security scanning of your code becomes possible inside your own infrastructure, so your project never leaves for an external cloud service, at a cost low enough to scan on every update rather than once a year.

What Antares Actually Does

Antares is not a general chat model or a coding assistant. It is an agent specialized in a single task called vulnerability localization. You give it a vulnerability type described as a CWE — the standard classification for software weakness categories, such as CWE-78 for OS command injection — along with a code repository, and it goes to work the way a security engineer does: it runs ordinary terminal commands like grep, find, and cat, reads the output, reasons about it, and narrows the search step by step. After a maximum of 15 commands it submits the list of files it believes contain the vulnerability, or declares the repository clean of that specific type.

The difference from traditional static analysis tools is that Antares relies on neither fixed rules nor a vector database; it actually explores the repository, forms a hypothesis, and verifies it. Cisco ships it with a ready command-line tool that operates on a read-only snapshot of your repository and returns results in human-readable, JSON, or SARIF format — meaning output that plugs straight into your existing security tooling and CI pipelines.

The Numbers That Matter to You

The models were measured on the Vulnerability Localization Benchmark: 500 real-world entries spanning 6 software ecosystems and 147 CWE categories. The published results:

• Performance: Antares-3B scored a File F1 of 0.223, beating every other open-weight model — including systems 250 times its size — and approaching GPT-5.5.
• Cost: It completed the full 500-task sweep for $0.82 in about 15 minutes on a single H100, roughly $0.0016 per task. For comparison: GLM-5.2 cost $12.50 in 50 minutes, and GPT-5.5 cost $141 across 5 hours — making Antares about 172 times cheaper.
• Sizes: the 350M version has a 32K-token context window and targets edge and IoT devices; the 1B version has a 128K window and runs on a workstation; the 3B version has a 128K window, runs on any single GPU, and generates more than 1,500 tokens per second.

One important note in reading these numbers: parameter count does not predict performance on this task at all. Llama-3.3-70B scored 0.012 and Qwen3.5-122B scored 0.091, while a model of just 3 billion parameters scored 0.223. Specialization beat size.

Why a Small Model Is the Right Choice Here

You might ask why not simply use a bigger, stronger model. The reason is threefold. First, privacy: security tools should run close to the code, and shipping your source to an external API opens supply-chain risk and moves your intellectual property outside your control. Second, cost: agents work across multiple turns, and every turn multiplies consumption, so what looks cheap on a single request becomes a heavy bill across thousands of scans. Third, speed: a scan that takes five hours cannot run on every push, while one that finishes in minutes becomes a natural part of your build pipeline.

What Antares Does Not Do — and This Matters

Honesty is more useful here than enthusiasm. A score of 0.223 is low in absolute terms, which means the model is wrong often and right sometimes. This is a triage tool, not a verdict tool: its job is to shrink the haystack from thousands of files down to a short list a human security engineer reviews. Cisco itself lists among the out-of-scope uses: security decision-making without human review, automated vulnerability remediation without testing, and final determination of compliance. Add to that a training data cutoff of 10 April 2025, meaning the model does not know about vulnerabilities discovered after that date.

What This Means for Your Business in Saudi Arabia

Three practical takeaways. The first concerns data protection: if you operate under the Personal Data Protection Law, every external service that touches your systems needs review, and a model running inside your own servers removes that question entirely. The second concerns cost: many mid-sized companies postpone security scanning because it is expensive and gets done once a year through an outside firm; at fractions of a halala per task, continuous scanning becomes an easy decision. The third concerns direction: small, specialized open models are improving quickly, and whoever builds systems today that can accept these tools into the development pipeline will benefit from every future improvement without rebuilding.

How We Approach This at Origami

As a technology company, we do not treat any new tool as a substitute for fundamentals. The basics still hold: disciplined code review, correct permission management, dependency updates, and security testing before launch. What tools like Antares add is a cheap early triage layer that runs inside continuous integration and warns your team before code reaches production. And when we build a system for a client, we design it to accept that layer later: an organized repository, an automated build pipeline, and security output in a standard format. That way every advance in this field becomes a direct gain for our client rather than a new project.

Sources: the official Cisco Antares project page at cisco-foundation-ai.github.io/antares, the official model card on Hugging Face, and the technical report "Antares: Foundation Models for Agentic Vulnerability Localization," Foundation AI, Cisco, 2026. All performance and cost figures are as published in Cisco's own evaluations.