On July 21, 2026, Cisco Foundation AI introduced Antares — compact security SLMs built for one expensive job: vulnerability localization (map a CWE or advisory to the files that likely contain the flaw).
Two sizes ship open-weight now: Antares-350M and Antares-1B. A 3B sibling is in the family but not the open drop. With them comes VLoc Bench — a 500-task agentic benchmark — and a clear thesis: small models with learned search can beat token-heavy generalists on this narrow task while keeping code on-prem.
TL;DR — What People Are Asking
| Question | Answer |
|---|---|
| What shipped? | Antares-350M + Antares-1B open weights |
| Who? | Cisco Foundation AI (Amin Karbasi et al.) |
| Task? | Vulnerability localization — ranked files + terminal trace |
| Bench? | VLoc Bench · 500 tasks · CWE-conditioned |
| Base models? | IBM Granite 4.0 checkpoints |
| Training? | SFT + GRPO (verifiable rewards on vuln repos) |
| 3B open? | Not in the July open drop |
| Replaces AppSec suite? | No — triage aid, not full toolchain |
The problem Antares targets
Connecting public vulnerability knowledge (advisories, CWEs, databases) to your repo is slow: large trees, noisy signals, evidence scattered across files and call paths.
Static analysis helps but dumps triage debt. Frontier coding models can reason about code but are not optimized for security investigation + terminal navigation + structured localization. Antares aims at that middle ground.
Cisco’s own framing of the dual pressure: agents write more code and get better at exploiting it — so cheap, continuous verification has to sit close to the loop. Small local models are compelling because proprietary code never has to leave the machine.
How Antares works (investigator loop)
| Step | Behavior |
|---|---|
| Start | Vulnerability description / CWE-style context |
| Search | Look for relevant patterns in the repo |
| Read | Inspect candidate files |
| Revise | Incorporate evidence; backtrack when a path is dead |
| Output | Ranked file list + terminal exploration trace |
Inspired by Cisco Foundation AI research that compact models can learn to search, reflect, revise, and backtrack — useful retrieval from strategy, not only from scale.
Workflows Cisco lists: CWE file localization, advisory-driven triage, augmenting SAST with model-driven exploration, early CI/CD review of likely vulnerable files, air-gapped / privacy-sensitive analysis.
VLoc Bench — why a new benchmark
| Benchmark family | Measures | Gap |
|---|---|---|
| SWE-Bench-class / CodeScout-style | Find code for a software issue | Not CWE/advisory localization |
| VLoc Bench | Navigate unfamiliar repos and recognize vuln patterns for CWE categories | Security-specific localization |
500 tasks. Agent must navigate efficiently and recognize vulnerability patterns. That is closer to how triage actually fails than “did the agent open the right file for a feature bug?”
Benchmark snapshot (vendor table)
Numbers below are from Cisco / model-card reporting on File F1 (higher is better). Always re-run on your harness.
| Model | Scale | File F1 (approx.) |
|---|---|---|
| GPT-5.5 (xhigh) | Frontier | 0.229 |
| Antares-3B (GRPO) | 3B | 0.223 |
| Antares-1B (GRPO) | 1B | 0.209 |
| GLM-5.2 | 753B-class | 0.186 |
| Gemini 3 Pro | Frontier | 0.152 |
| Antares-350M (GRPO) | 350M | 0.135 |
| Granite 4.0 1B (base) | 1B | ~0.000 |
explainx.ai read:
- Post-training supplies essentially all the localization skill — base Granite is near zero under the same protocol.
- GRPO improves File F1 and cuts run-to-run variance (Cisco reports double-digit % lifts) — important for CI.
- Cost/runtime claims: full 500-task sweep on the order of ~15 minutes on one H100 / <$0.002 per task amortized in the technical report — orders of magnitude under some frontier API sweeps. Verify yourself.
Where Antares sits in the defender stack
Cisco ties Antares to a broader Foundation AI story:
| Piece | Role |
|---|---|
| Foundry Security Spec | Model-agnostic blueprint for agentic security eval systems |
| CodeGuard | Secure-by-default rules/skills for coding agents |
| Antares | Compact models + benchmark for repo-level localization |
Compare to other Jul 2026 defender products we cover:
| Product | Angle |
|---|---|
| OpenAI Daybreak / Codex Security | Agentic defense workflows, gated cyber models |
| Claude Security | Anthropic public-beta security scanning |
| Antares | Open-weight, local, localization-first SLMs |
Different trust boundaries: Antares optimizes for keeping source inside your network; frontier cyber products optimize for capability under access programs.
Practical limits (read before you CI-gate)
- Localization ≠ exploit validation — ranked files still need human review.
- Not SCA / secrets / DAST / containers — keep the rest of AppSec.
- Antares-3B (strongest sibling in the table) is not the open July drop.
- Phase B false-alarm rates and production FP economics need your own measurement.
- Agent terminal navigation in CI needs the same sandbox discipline as any coding agent.
Who should try it this week
| Team | Move |
|---|---|
| AppSec with air-gap rules | Pull Antares-1B; run VLoc-style tasks on an internal mirror |
| University / public-sector security | Prefer 350M/1B local over sending repos to APIs |
| Platform eng | Wire ranked-file output into existing triage queues — not as auto-close |
| Red/blue researchers | Read the technical report; compare against your CodeScout/SWE harnesses |
Related on explainx.ai
- OpenAI Daybreak — Codex Security cyber defense
- Claude Security Scan Fix public beta
- Claude Mythos / Glasswing cybersecurity
- Hugging Face autonomous agent breach
- AI cyber guardrails blocking US defenders
- Sakana Fugu — cyber orchestration
- What is an agent harness?
- Agent skills security threat verification
- Is Claude Cowork safe?
- Bonsai 27B — dense local models (different task)
Sources: Cisco — Introducing Antares · Cisco Foundation AI technical report / Hugging Face model cards (Jul 21, 2026)
Benchmarks and cost figures reflect Cisco Foundation AI’s July 21, 2026 materials and model cards. Security tooling results vary by codebase and harness — do not treat File F1 as a substitute for your vulnerability management SLAs.