Antares: Open Weight AI Models for Vulnerability Localization
16 hours ago
- Cisco releases Antares, a family of open-weight small language models (SLMs) for security vulnerability localization, available on Hugging Face.
- Antares models (350M and 1B parameters) outperform larger closed- and open-weight models on the new Vulnerability Localization Benchmark at lower cost.
- The models are compact enough to run locally, enabling sensitive codebases to stay on-premises without cloud transmission.
- Antares mimics human analyst search patterns, iteratively navigating code to identify vulnerable files based on CWE descriptions.
- The release includes a new 500-task benchmark tailored for security-specific localization, filling a gap in existing coding benchmarks.
- Cisco's broader initiative includes Foundry Security Spec, CodeGuard, and the benchmark to foster practical, trustworthy AI tools for cybersecurity.