Cisco Antares: Open-Weight AI Models That Find Code Vulnerabilities Cheaply

What Cisco actually released

Cisco has open-sourced two small language models, Antares-350M and Antares-1B, built specifically for vulnerability hunting in software repositories.

These are not positioned as general coding assistants that happen to know some security. They appear designed to act more like repo investigators: search the codebase, inspect likely files, change course when a lead goes cold, and narrow in on where a known vulnerability may actually live.

Cisco is also keeping a more capable Antares-3B model out of broad public release for now, with plans to use it in its own security products.

Why this matters

Repository security scanning has an awkward economics problem. Teams need to rescan often because software changes often, but running large models repeatedly across large codebases can turn into a pricey habit.

Cisco’s argument is that small, specialized models are more practical for this job. In the company’s testing, Antares scanned 500 repositories in about 15 minutes for under $1, versus roughly five hours and more than $100 for GPT-5.5.

That kind of gap matters for one reason above all: repeatability. A scan that is cheap enough to run often is more useful than a very smart scan you hesitate to trigger.

The privacy angle is doing real work here

For many teams, “local” is the headline.

Because the smaller Antares models are designed to run on-premise or inside an organization’s own environment, companies may be able to keep sensitive source code in-house rather than sending it to a third-party AI provider. That changes the appeal quite a bit for regulated industries, internal security teams, and anyone who flinches at uploading a crown-jewel repo to the cloud.

In other words: less exposure, fewer procurement headaches, and fewer meetings where someone says “Can legal look at this?”

Antares is specialized on purpose

Cisco says it trained Antares to think less like a chatbot and more like a security analyst following breadcrumbs.

That distinction matters. Vulnerability discovery in real repositories is not just about answering “Is this code bad?” It often means exploring a huge codebase, tracing context, checking related files, and identifying the few lines that actually matter.

A general-purpose model may still help, but Cisco is betting that a smaller model trained for this exact workflow can punch above its weight. Based on the available context, that is the real product strategy here: narrower scope, lower cost, more operational fit.

Open-weight, but not a free-for-all

Cisco plans to make the models available on Hugging Face to users it verifies as cyber defenders.

Open-weight security tools can be useful for defenders and attractive to attackers, so Cisco appears to be trying to thread the needle between openness and misuse control. The company also says it worked with U.S. government agencies on safety and release decisions.

This is not full public download-and-go. It is more like “open, with a bouncer.”

The benchmark question

Cisco says Antares performed similarly to much larger models, including GPT-5.5 and GLM-5.2, on its own benchmark for identifying vulnerable files in real-world repositories.

That is interesting, with one obvious caveat: company-created benchmarks are helpful, but they are still company-created benchmarks. The broader test will be whether security teams can reproduce useful results on their own repos, workflows, and threat models.

Still, the claim points to a bigger trend worth watching: small models are getting more compelling when the task is narrow, repetitive, and expensive at scale.

This fits a wider shift in AI security tooling

Cisco is not alone here. The market is filling up with AI security tools that are less interested in being universal assistants and more interested in doing one painful job well.

That includes tools that review code from an attacker’s perspective, triage likely weak spots, or continuously scan repositories for known bug patterns. The shared theme is practical specialization. Not “one model to rule them all,” but “one model to stop wasting time on this specific task.”

For buyers, that is usually a healthier direction.

Who should care

Antares looks most relevant for:

• Security teams scanning large, active codebases
• Enterprises that want on-premise or local AI options
• Organizations wary of sending proprietary code to external providers
• Teams trying to reduce the cost of repeated vulnerability scans

It may be less compelling if your workflow depends on broad code generation, debugging help, or a single model that handles every developer task. Antares seems built for one lane, not all lanes.

What to watch next

There are three practical things to monitor:

• Whether defenders find the verified-access model workable
• Whether independent testing supports Cisco’s performance and cost claims
• Whether more vendors ship small, local, security-specific models instead of defaulting to giant general-purpose ones

Cisco is also exploring an industry consortium around open AI security tools, which suggests this release may be less of a one-off and more of a positioning move.

The useful takeaway

If you care about secure code analysis, Antares is a good reminder that bigger is not always better. For repository scanning, the winning setup may be the model you can afford to run constantly, keep local, and trust with sensitive code—not the one with the flashiest demo.