Antares scans five hundred code repositories for under a dollar. A model a fraction of the size of the frontier systems it's benchmarked against is finding nearly as many of the right vulnerable files, at a cost measured in fractions of a cent instead of dollars per task.

Cisco released Antares-350M and Antares-1B on Hugging Face this week, two small language models Cisco Foundation AI trained for one job: finding which files in a codebase contain a known vulnerability, given nothing more than a vulnerability category. The weights go through a request process on Hugging Face; the benchmark and evaluation code behind these numbers are fully open on GitHub.

The Cost Case Holds Up

Software teams already know the expensive part of vulnerability response. A public advisory names a weakness class. Finding which files in a million-line repository contain it requires an analyst to trace call paths, compare naming conventions, and rule out false leads by hand. Antares automates that search. Given a vulnerability category and read-only terminal access to a repository, it searches, reads candidate files, and narrows toward a ranked list of likely culprits, the way an investigator works a lead rather than the way a chatbot answers a question.

On Cisco's new Vulnerability Localization Benchmark, Antares-1B reaches a File F1 score of 0.209, ahead of Z.ai's 753 billion parameter GLM-5.2 at 0.186 (MarkTechPost, 2026). OpenAI's GPT-5.5 still leads at 0.229, and costs accordingly: roughly $141 to sweep the same 500 repositories that Antares clears for under a dollar on a single H100 (MarkTechPost, 2026; Axios, 2026). Traditional static analysis tools trail all of them. Semgrep scores 0.086, CodeQL 0.023 (MarkTechPost, 2026).

That gap is the real story, and it holds regardless of what happens next in this piece. A small, purpose-built model is finding more of the right files than a general-purpose model hundreds of times its size, faster, for less money, and it can run inside a customer's own environment instead of sending proprietary source code to a third party. For a university security team or a public-sector agency priced out of frontier-model token budgets, that difference decides whether continuous scanning happens at all.

"You really don't need a private jet to go to your corner store." — DJ Sampath, Cisco senior vice president and general manager of AI software and platform (Axios, 2026)

Built to Do One Narrow Thing

Antares localizes files tied to already-known vulnerability classes. It does not assess exploitability, generate exploit code, or attempt zero-day discovery, and Cisco's own model card says so directly: the model identifies which files are likely affected, not why or how to exploit them (Hugging Face, 2026). Cisco also worked with United States government agencies on the release, and access to the weights runs through a request process rather than an instant download.

Cisco has also not released Antares-3B, the version that scores closest to GPT-5.5 at 0.223 File F1. Cisco says that model will go into its own security products instead, which means the model Cisco benchmarks against GPT-5.5 in its own materials is not the one available to try today.

A Complementary Tool, Five Days Apart

Capital One released its own agentic security tool five days earlier, on July 16. VulnHunter is not a model of its own. It is a Hunt, Fix, Verify harness built around existing foundation models, currently Anthropic's Claude Opus 4.8 through Claude Code, that reasons from an attacker's perspective, maps out exploitable paths in source code, and proposes fixes. Capital One put the harness on GitHub under an Apache 2.0 license.

The two tools sit at different points in the same pipeline more than they compete. Antares is the underlying model, narrow and cheap, built to answer one question: which files likely contain a known vulnerability. VulnHunter is the broader workflow wrapped around a general model, built to investigate a candidate file further, trace whether it is actually exploitable, and verify a fix. A team could plausibly run Antares first to narrow a repository down, then hand the result to something like VulnHunter for the deeper investigation.

What the Benchmark Doesn't Show Yet

Cisco's benchmark runs two phases. Phase A hands the model a vulnerable snapshot and scores how many of the right files it finds. Phase B hands the model a patched snapshot, the fix already applied, and checks whether it still raises an alarm on clean code.

Cisco published Phase A results for every model it tested. Phase B results are not out yet for any of them.

That gap matters more than the headline F1 score for anyone planning to gate a continuous integration pipeline on this model's output. A vulnerability scanner that flags clean code at scale trains developers to ignore it within a month. Without a published true negative rate, a security team adopting Antares today is running its own experiment on false-alarm tolerance, with no benchmark yet to check it against.

Where Antares Sits in the Stack

Antares sits in a different layer of Cisco's security buildout than the identity work covered here this spring. Astrix, Galileo, and WideField Security all fed into Splunk's data plane, assembling a discovery, evaluation, and session-evidence stack for non-human identities in eight weeks. Antares operates upstream of that, at the point where a vulnerability advisory meets a codebase, before an agent or a credential enters the picture at all.

The pattern fits Cisco's other recent moves too. Cisco has generally kept its most capable engines inside its own commercial products while open-sourcing the components built to earn trust and adoption around them. Antares-3B moving into Cisco's own security line while the smaller models validate in the open follows that same logic.

The practical question for a buyer is whether the smaller models perform well enough on their own, without Antares-3B and without published false-alarm data, to be worth building a workflow around this year.