# Cisco's Antares Models Target the Gap Between "Too Expensive" and "Too Noisy" in Vulnerability Triage
Every time a CVE drops, the same silent race begins inside security teams at companies that ship software. Someone has to answer the question: *does our code have this?* Not the generic "are we running a vulnerable version" question — the harder one. Does the actual vulnerability pattern exist somewhere in the codebase? Which files? Which functions?
Most teams answer this badly, slowly, or not at all until something breaks. Cisco is betting a pair of small open-weight models can change that calculus.
## The Real Problem: Two Bad Options and a Clock That's Already Running
The current state of code vulnerability triage is a mess that doesn't get enough attention. When a critical CVE lands — especially something like a memory corruption issue or an authentication bypass tied to a specific code pattern — defenders are essentially in a race. Attackers already have the CVE details. They're looking for the same vulnerable pattern you are, just from the other side.
Large frontier models like GPT-5.5 can actually do this well. They understand code context, they can search meaningfully across unfamiliar repositories, and their false positive rates are manageable. But at the volume a real security team generates these queries — every significant CVE against every internal codebase — the cost compounds fast. Add data sovereignty concerns and suddenly half the regulated industries (finance, healthcare, defense contractors) can't use them at all. You can't send your proprietary source code to an external API when it's subject to ITAR or GDPR or any number of sector-specific frameworks.
So teams fall back on smaller open-weight general language models. They stay in-house, the cost drops dramatically, and the data never leaves. The trade-off is brutal: false positive rates climb sharply. Now instead of narrowing down where to look, you've generated a list that contains everything. Security analysts end up manually reviewing files that aren't vulnerable — which defeats the purpose entirely.
Cisco's Antares models are designed to live in the gap between those two options.
## How It Actually Works
Antares-350M and Antares-1B (the naming reflects parameter count) are small language models, not general-purpose code assistants. They're trained specifically for one task: given a vulnerability description, navigate an unfamiliar codebase and surface the files most likely to contain a matching pattern.
The search process Cisco describes is worth paying attention to. Rather than running a static grep-style pattern match, the model works iteratively — starting from a vulnerability description, finding candidate files, reading them, updating its confidence, changing direction when a path doesn't pan out, and converging on a ranked list. Cisco compares this to how a human investigator would actually work through a repository.
This matters because many vulnerabilities aren't a single easily-identifiable pattern. A use-after-free or an integer overflow might manifest differently depending on the language version, the surrounding code structure, or how abstractions were layered on top. A model that can follow the reasoning chain — rather than just scanning for literal strings — should produce fewer false positives than a general model that doesn't have this task burned into its architecture.
The output is a ranked list of source files. Not a verdict, not a fix — a starting point for human triage. Cisco is explicit that this is a first-stage assist, not a replacement for expert review.
## Benchmarks and the Strategic Move Nobody's Talking About
Alongside the models, Cisco released the Vulnerability Localization Benchmark — 500 entries mapped against CWE categories, designed to test exactly how well a model can navigate unfamiliar codebases and recognize vulnerability patterns simultaneously.
On this benchmark, Antares comes out well: 172x cheaper than GPT-5.5, 15.2x cheaper than Z.ai's GLM-5.2, with competitive accuracy.
Here's what's worth noting about releasing your own benchmark alongside your own models: Cisco has defined the terms of comparison in a domain that didn't have a standard measurement framework before today. That's not inherently sinister — creating benchmarks is how the field matures — but every consumer of this data should understand that the benchmark was designed by the same team that built the models being tested on it. Independent replication of these results will be important before anyone anchors critical decisions around vulnerability triage to Antares performance numbers alone.
That said, the benchmark itself looks rigorous on paper. 500 entries is a real dataset, CWE-category testing reflects how practitioners actually think about vulnerability classes, and requiring simultaneous codebase navigation plus pattern recognition is a meaningful combined test. If external researchers validate the methodology, this could become a useful industry standard.
## Who This Actually Helps
Cisco is pitching Antares specifically at universities, research institutions, nonprofits, and public-sector teams. These organizations frequently maintain important software — open-source infrastructure, government systems, educational platforms — with security resources that don't match the scale of their attack surface. A state government agency running a public benefits portal can't afford a frontier LLM contract, but they also can't afford to miss a critical vulnerability in their codebase.
The open-weight approach also matters for any organization that can't accept data egress risk. Source code is often the most sensitive proprietary asset a company owns. Running vulnerability triage against an external API is a real risk calculus, not a paranoid edge case.
The models are available on Hugging Face now. The activation energy for a small security team to try this is low.
## HackWire Analysis
The announcement lands at an interesting moment in the AI-for-security space. The dominant narrative for the past two years has been that bigger models mean better security tooling — that frontier LLMs are the only serious option for anything requiring genuine reasoning over complex code. Cisco is pushing back on that directly, and they're not the only ones.
What's happening here is a segmentation of the AI security market that mirrors what happened in EDR: enterprise solutions optimized for capability and compliance, and a second tier of purpose-built, affordable tools for the organizations that actually run most of the internet's infrastructure but can't afford the enterprise tier.
The false positive problem with general open-weight models isn't talked about enough in coverage of AI security tooling. Security analysts who've tried using Llama or Mistral variants for vulnerability research know the experience: you get hits everywhere, and you spend more time eliminating noise than you saved on the search. A model that genuinely reduces that noise — even moderately — has real operational value, not just benchmark value.
The comparison that comes to mind is the rise of purpose-built SAST tools versus trying to use general static analyzers for security findings. Semgrep, CodeQL, and similar tools succeeded not because they were smarter than everything else, but because they were tuned for the specific problem. Antares is attempting the same trick in the LLM space.
Whether Cisco can prove the false positive claims in practice — outside a benchmark they designed — is the open question. The next six months of external evaluation will be more informative than today's announcement. For now, the right move for teams with constrained budgets is to run Antares against repositories where they already know the ground truth and see whether the ranked output matches what they'd find through manual review.
Independent validation is the only kind that counts.
— HackWire Editorial
## Related Coverage