July was an Antares month, with a parallel thread on smarter evaluation. Here’s what we shipped.
**Antares (technical report)****. **A family of compact, open-weight models (350M, 1B, and 3B) built for one job: agentic vulnerability localization. Given just a CWE description and read-only access to a repository, Antares searches the codebase and pinpoints the files where the vulnerability lives. The headline: the 3B model rivals frontier systems like GPT-5.5 while running locally at a tiny fraction of the cost, and it outperforms open models many times its size. The thesis behind it — for security, a small model trained for the actual task beats a giant general-purpose one, and inference stays in-house so your source never leaves the machine. Antares-350M and Antares-1B are publicly released. Within a month, they have reached 29,800 downloads on Hugging Face.
**Antares Quickstart (cookbook)****. **The practical companion: a CLI that turns the model into something you can run in ~15 minutes. It profiles a repo, checks for relevant weaknesses, and produces reports (including SARIF for GitHub Code Scanning), with privacy-preserving defaults and CI gating. This is the bridge from “impressive benchmark” to “thing you gate every push with.”
**Antares in the Wild (blog, Jul 28)****. **A week after release, the team ran the unmodified models on two benchmarks built by other people: a Berkeley cyber-capability suite and a community rebuild of Devin’s security eval. The finding: localization holds up off-distribution, matching hosted agents that cost orders of magnitude more per run. The capability generalizes, not just fits one benchmark.
**New Reasoning Benchmarks in FAITH (blog, Jul 17)****. **Two additions to our evaluation hub, CTI-Reasoning and CWE-Prediction, built to measure whether a model can reason about security rather than recite memorized facts: using threat-intelligence analysis tasks and vulnerability data too recent to have been trained on. Both are deliberately hard and unsaturated, leaving headroom even for frontier models.
**Black Hat USA 2026 (Aug 4–6). **The Foundation AI team brought Antares and our research on fully automated prompt optimization to Las Vegas, presenting at Cisco booth sessions and demonstrating the work at the AI Kiosk. We gathered valuable feedback from Antares users while introducing the model family to more security researchers, penetration testers, and security engineers. These conversations reinforced Antares’s core idea: compact, task-specific models can deliver practical security value while keeping sensitive code local. With agent evaluation and improvement emerging as major themes across the conference, the event also underscored the potential of our open-source prompt optimization project to help autonomous agents continuously evolve. The interest in prompt optimization further demonstrated how open-sourcing this capability can give the community a practical way to build agents that learn from evaluation results and improve with less manual tuning.