Cisco Foundation AI Releases Antares: 350M and 1B Open-Weight Models That Localize Known Vulnerabilities Inside Real Codebases
Cisco Foundation AI released Antares, a family of open-weight security small language models (SLMs) specifically designed for vulnerability localization within software repositories. The Antares-1B model achieves a File F1 score of 0.209, outperforming significantly larger models like GLM-5.2 (753B parameters) and rivaling GPT-5.5, demonstrating that task-specific training outweighs raw parameter scale. Performance gains are driven almost entirely by post-training techniques, with Supervised Fin
Analysis
TL;DR
- Cisco Foundation AI released Antares, a family of open-weight security small language models (SLMs) specifically designed for vulnerability localization within software repositories.
- The Antares-1B model achieves a File F1 score of 0.209, outperforming significantly larger models like GLM-5.2 (753B parameters) and rivaling GPT-5.5, demonstrating that task-specific training outweighs raw parameter scale.
- Performance gains are driven almost entirely by post-training techniques, with Supervised Fine-Tuning (SFT) providing the foundational capability and Group Relative Policy Optimization (GRPO) adding further accuracy while drastically reducing output variance.
- The release includes VLoc Bench, a rigorous 500-task agentic evaluation benchmark covering real-world repositories across six major ecosystems, establishing a new standard for measuring security-focused LLM performance.
- Antares operates via a constrained agentic loop using read-only terminal commands, offering a cost-effective alternative to static analysis tools and massive proprietary models for initial security triage.
Why It Matters
This release challenges the prevailing assumption that larger models are always superior for specialized technical tasks, proving that focused fine-tuning on small architectures can surpass billion-parameter generalist models in niche domains like cybersecurity. For AI practitioners and security engineers, it highlights the critical importance of high-quality, domain-specific post-training data and reinforcement learning techniques over mere model scaling. Furthermore, the availability of open-weight models and a standardized benchmark enables the community to build more efficient, transparent, and cost-effective security toolchains that integrate seamlessly into continuous integration pipelines.
Technical Details
- Architecture & Training: Antares models (350M, 1B, 3B) are decoder-only transformers initialized from IBM Granite 4.0 checkpoints, utilizing grouped-query attention, SwiGLU MLPs, RMSNorm, and RoPE. The 350M and 1B variants are open-weight under Apache 2.0.
- Agentic Workflow: The models function within a constrained agent loop, issuing up to 15 read-only terminal commands in a Docker sandbox to explore repositories. They receive only CWE descriptions without advisory hints, requiring them to autonomously locate vulnerable files.
- Post-Training Pipeline: Capabilities are derived from SFT (using traces from GPT-OSS-120B) followed by GRPO. GRPO uses verifiable, programmatic rewards based on trajectory text, eliminating the need for learned reward models and significantly cutting run-to-run variance by 42-65%.
- Benchmarking (VLoc Bench): Evaluated on 500 tasks from 290 real-world repositories across npm, pip, Maven, Go, Rust, and Composer. Phase A measures File F1 on vulnerable snapshots, while Phase B assesses True Negative Rates on patched snapshots to detect false alarms.
- Performance Metrics: Antares-1B achieved 0.209 File F1 and highest recall (0.224). It outperformed static analysis tools like Semgrep (0.086) and CodeQL (0.023). The base Granite checkpoints scored near zero (~0.000), confirming post-training supplies nearly all utility.
Industry Insight
- Shift from Scale to Specialization: Organizations should prioritize developing or fine-tuning smaller, domain-specific models for high-volume, narrow tasks like security triage rather than relying solely on expensive, large-scale generalist models. This approach offers better cost-efficiency and potentially higher accuracy for specific workflows.
- Importance of Deterministic Evaluation: The significant reduction in variance provided by GRPO suggests that reliability and consistency are as crucial as peak performance metrics for deploying AI in production security environments. Teams should evaluate models not just on average scores but on stability across multiple runs.
- Integration into DevSecOps: Since Antares is designed to complement, not replace, existing toolchains, developers should view these SLMs as intelligent filters that reduce manual effort in identifying potential vulnerabilities, allowing human experts to focus on complex remediation rather than initial discovery.
Disclaimer: The above content is generated by AI and is for reference only.