Cisco's Antares AI Helps Code Reviewers Find Vulnerabilities Faster
Cisco's Antares models narrow a vast codebase down to a ranked shortlist of suspect files, giving security teams a faster path into triage without replacing the humans who do it.
Key points
- Cisco launched Antares, a family of open-weight AI models, designed for repository-scale vulnerability localization.
- The models accept a Common Weakness Enumeration (CWE) description and return a ranked list of files most likely to contain that class of flaw.
- Three model sizes ship: 350 million, 1 billion, and 3 billion parameters.
- Cisco says the 3-billion-parameter model approaches GPT-5.5 on its internal vulnerability localization benchmark.
- All three sizes are available as open-weight models on Hugging Face.
Cisco has released Antares, a family of AI models built to shrink a massive software repository down to a focused list of files worth a human's attention. Feed it a CWE description, the standard taxonomy of software weakness classes, and it returns a ranked shortlist. No CVE match, no patch, no verdict on severity. Just a triage starting point.
We covered Cisco's earlier work in this area on 21 July; Antares has since been formally detailed with benchmark numbers attached.
How does Antares differ from existing tools?
It does not work the way Semgrep or CodeQL do. Those tools match code against predefined rules. Antares instead works as what Cisco calls an exploration agent, adjusting its path through the repository as it gathers evidence rather than firing a fixed ruleset at every file. Amin Karbasi, Cisco Foundation AI's chief scientist, wrote in a blog post that the output is a ranked file list plus a trace of how the model reached it. That trace matters: reviewers can see the reasoning, not just the answer.
Supriti Vijay, a Cisco AI researcher, said via email that the aim is to "reduce fatigue and workload by helping them triage an issue earlier and focus their investigation on the most relevant parts of the codebase." The model does not replace the engineer's call on whether something is actually exploitable or how bad it is.
Should you worry about running it on proprietary code?
The CLI supports local inference, which means proprietary code stays inside your own environment rather than touching an external API. For any team already nervous about sending source code to a third-party service, that matters. The three model sizes map roughly to deployment context: the 350-million-parameter version for constrained environments, the 1-billion for a laptop or workstation, the 3-billion for a single GPU.
The benchmark claim deserves a sceptical read. Cisco compared Antares-3B against GPT-5.5 on Cisco's own internal Vloc benchmark. That's not independent validation, and the comparison models from Google, OpenAI, and Meta were larger open-weight versions, not the frontier proprietary ones. Specialised beats general on narrow tasks is a real phenomenon, but "approaches GPT-5.5" on a vendor-designed test is a headline number, not a deployment guarantee.
The honest question for any team evaluating this is how often repository-wide scans fit into an existing workflow and whether the reduction in file-review effort actually shows up as time saved. Antares finds where to look. Everything after that is still on you.
