Cisco Foundation AI just open-sourced Antares, a family of small language models that do one thing: find the vulnerable files a security advisory is talking about. The 1B-parameter version scores a 0.209 File F1 on the new VLoc Bench, landing within striking distance of GPT-5.5’s 0.229 — and notably ahead of a 753B open-weight model that manages just 0.186.
It’s a capability cliff, not a smooth scaling curve. The untrained base checkpoints all score near zero. Supervised fine-tuning on a corpus that’s 71.5% cybersecurity reasoning does the heavy lifting, pushing the 1B model to 0.188. Then GRPO reinforcement learning — using programmatic rewards, no learned reward model — adds another 11%, while also slashing run-to-run variance by up to 65%. The teacher for all reasoning traces is a single model, GPT-OSS-120B, to keep the distribution clean.
The agent loop is constrained and realistic. The model gets a CWE category description and a Docker sandbox with no network access. It has 15 terminal calls to explore the repo and must return a ranked list of vulnerable files. VLoc Bench itself is 500 tasks drawn from real GitHub Security Advisories across npm, pip, Maven, Go, Rust, and Composer. Ground truth comes straight from the actual security patch.
Cisco is explicit that Antares doesn’t replace your SAST, dependency scanners, or threat modeling. It targets the expensive first triage step where a developer is staring at an unfamiliar repo trying to map a public CVE to actual code. And the release includes not just the 350M and 1B models under Apache 2.0, but the full VLoc Bench evaluation suite. That’s the kind of concrete, reproducible security research the field needs more of.
