Antares: Foundation Models for Agentic Vulnerability Localization

How can compact language models be trained to autonomously navigate large, unfamiliar software repositories to identify the specific files containing security vulnerabilities, given only a high-level vulnerability description?

The authors introduce Antares, a family of models (350M, 1B, and 3B parameters) based on IBM Granite checkpoints. The training pipeline consists of two stages:

1. Supervised Fine-Tuning (SFT): The models are trained on a corpus of cybersecurity reasoning, repository exploration, and terminal interaction data.
2. Reinforcement Learning (RL): The models undergo Group Relative Policy Optimization (GRPO) using verifiable rewards based on successful file-level localization within a constrained terminal environment.

During inference, Antares acts as an agent, iteratively issuing read-only terminal commands to explore a codebase, inspect files, and refine its search until it submits the location of the vulnerability or exhausts its command budget.

Antares-3B demonstrates performance that approaches significantly larger, closed-source frontier models (like GPT-5.5) while outperforming open-weight models that are over 200 times larger. The reinforcement learning stage is critical, as it induces an emergent search-verify-refine strategy that is more effective than standard repository browsing. Furthermore, the models are highly efficient, capable of completing a 500-task evaluation sweep in approximately 15 minutes on a single H100 GPU, making them suitable for local, low-cost, and privacy-preserving security workflows.

Vulnerability localization is a bottleneck in security remediation. By enabling high-performance, agentic localization on compact models, Antares allows organizations to run security agents locally within their own trust boundaries, avoiding the privacy risks and costs associated with sending proprietary code to external APIs.