Sunday, November 15, 2026

What is Cisco's Antares?

Cisco's Antares is a family of compact, open-weight language models purpose-built for agentic vulnerability localization in codebases, operating alongside the Foundry Security Spec. [1, 2]


Antares Model Family Specs
Antares models receive a Common Weakness Enumeration (CWE) description and terminal access to a repository, using multi-turn reasoning and tool use to shortlist vulnerable source files. [1]
  • Antares-350M
    • Base Model: IBM Granite 4.0 350M
    • Context Window: 32K tokens
    • Target Deployment: Mobile, IoT, and edge devices
    • Performance: File F1 score of 0.135 [1]
  • Antares-1B
    • Base Model: IBM Granite 4.0 1B (40 layers, hidden dim 2048, 16 attention heads, GQA, SwiGLU, RoPE)
    • Context Window: 128K tokens
    • Target Deployment: Laptops and workstations
    • Performance: File F1 score of 0.209 (highest recall at 0.224) [1, 2]
  • Antares-3B
    • Base Model: IBM Granite 4.0 Micro 3B
    • Context Window: 128K tokens
    • Target Deployment: Single GPU setups
    • Performance: File F1 score of 0.223 (approaching GPT-5.5 performance on benchmarks) [1]

The Integrated Security Ecosystem
Antares is part of a three-part framework released by Cisco Foundation AI: [1]
  • Antares (Localize): Takes a security advisory/CWE and outputs a ranked file shortlist of affected source code. [1]
  • Foundry Security Spec (Evaluate): A model-agnostic blueprint and harness providing clear roles, guardrails, evidence-gated triage, and reviewable outputs. [1, 2]
  • CodeGuard (Prevent): Contributes secure-by-default rules and skills to guide AI coding agents and block weaknesses from being reintroduced in the IDE. [1, 2]

No comments:

Post a Comment