In a dizzyingly short time, large-language AI models have revolutionized how we think, work, and solve problems.
But these giant-scale models bring challenges as well — particularly related to security and cost.
That’s because the same large-language capabilities used by defenders to pinpoint vulnerabilities deep within code bases can be exploited by bad actors. And for some organizations, using large-language models (LLMs) to pinpoint vulnerabilities ahead of the attackers can be too expensive, while exposing their data to potential dangers in the cloud.
To meet these challenges, Cisco introduced Antares, a family of specialized small-language model (SLM) security tools, the first two of which are now accessible as “open weight” releases on Hugging Face. The goal is to make secure AI available to many more organizations, including at the non-enterprise level.
To learn more, we spoke with Amin Karbasi, VP and chief AI scientist at Cisco.
Thank you for meeting with us, Amin! Let’s start with a quick overview of the security challenges that led to the Antares models.
Thank you, Kevin. The name Antares captures the problem that we are facing very nicely. Antares is an enormous red star, but because it is far away, it is not visible due to our sun. The sun dominates our sky, right? And vulnerabilities are similar. There are maybe only a few vulnerable lines hidden within millions of lines of code. But their impact can be enormous.
So, Antares pinpoints those vulnerabilities.
Yes, this is a new tool for the security analysts. When they're bombarded with these giant repositories, Antares narrows the search for them to very specific files. The internal code and repositories are usually very, very large. Signals are very noisy, and the relevant evidence might be spread across multiple files and many, many paths. Antares is trained specifically to explore these repositories, read candidate files, revise its search based on the evidence that it collects, and then return a ranked list of files for the analysts and investigators.
In the age of LLM AI, just about any organization could be vulnerable. But who would benefit most from Antares?
As you say, this challenges every organization that develops or maintains software. But the burden is especially significant for universities, nonprofits, public-sector institutions, and research organizations with smaller security teams. They face the same threats as large entities, but they may not have the people, infrastructure, or the budget to run expensive forensics. Or to use the frontier models. Others may have security, privacy compliance requirements, preventing them from sharing their code with third parties, in the cloud.
How does Antares compare to or complement the large frontier models?
Antares is very complementary to frontier models. Those frontier models are extremely capable, and they will remain valuable for broader tasks, broad reasoning for complex situations, validation, and remediation. Antares is aiming to solve a very narrow task, which is navigating a repository and basically localizing vulnerabilities very efficiently.
So, that ability to run locally is an important distinction. Yes, it’s a key strategic advantage of Antares. Source code reveals intellectual property, product architecture, security controls, and potential sensitive customer or operational information. So, many organizations have legitimate concerns about how they share data with third parties. That’s not a criticism of frontier providers. It is a question of control.
Tell us about the three different versions of Antares, and the users and use cases for which each could apply.
We actually trained a class of models. Starting with a very small one, which is 350 billion parameters. That is designed for highly constrained environments, and it can actually run on browsers and smart phones.
The 1B is maybe the primary practical release. It balances performance, costs, and deployment flexibility. It is suitable for laptops and workstations. Then there is the bigger 3B, the most capable model that approaches frontier-level performance, while remaining very small. It runs on a single GPU. We have the model in house, but we are gating it, to be very responsible about the release because it's a powerful model.
Powerful in the sense that it could fall into the wrong hands?
Yes, powerful in that it is getting very close to the frontier models’ capabilities in vulnerability localization. So, we want to partner with our customers who need these models to make sure that they are the ones who get it and use it responsibly.
Cisco has long been committed to democratizing security and connectivity. How does Antares, with its open-weight release, align with that tradition?
Security should not be a luxury available only to organizations with frontier-AI scale budgets. And so, Antares, which is open weight, offers much easier access to intelligence and control. With coding agents, we are producing more and more code — and more and more vulnerabilities. So, finding those vulnerabilities is not a one-time task. You have to do it over and over. So, having access to a local model that is very cheap is extremely important. We are offering models that are much faster and much more economical than the frontier models.
How does Antares integrate with other Cisco offerings?
Again, it has become extremely difficult for cybersecurity analysts to do their job because they don't know where to look. So, we are giving them an important new tool. But it’s in combination with other tools that we have built at Cisco, like Foundry Security Spec and CodeGuard, which are helping to tailor frontier models to fix software flaws and introduce fewer vulnerabilities when writing code.
And that aligns with Cisco’s platform strategy.
Yes, eventually everything will be connected through Cisco Cloud Control. It’s going to be one platform that integrates it all.
What do you see for the future? Can we expect more tailored, security-specific models?
I think it will be a combination of frontier general-purpose models, smaller specialized models, and the orchestration to route work to the right intelligence. I don't think there is just one kind of frontier. A large general-purpose model is one frontier; a compact model that solves a very specific problem equally or better than frontier models that could be a thousand times bigger, is another frontier.
How do you view Cisco’s role moving forward?
Cisco has a distinctive role because we bring together deep, deep expertise in security, networking, observability, and infrastructure. Cisco is building the critical infrastructure for AI. It is also by definition a security company because almost 80% of the internet is running through Cisco switches. So, we are bringing all this expertise to the table.
Our goal is to create specialized models, open-benchmark practical systems that the community can evaluate, integrate, and improve. Cisco embodies the idea that intelligence that is specialized, that is affordable, locally deployable, and available for the broader security community is something to build upon. Antares is just the beginning step in that direction.
