July was an Antares month, with a parallel thread on smarter analysis. Right here’s what we shipped.
Antares (technical report).A household of compact, open-weight fashions (350M, 1B, and 3B) constructed for one job: agentic vulnerability localization. Given only a CWE description and read-only entry to a repository, Antares searches the codebase and pinpoints the recordsdata the place the vulnerability lives. The headline: the 3B mannequin rivals frontier techniques like GPT-5.5 whereas operating domestically at a tiny fraction of the associated fee, and it outperforms open fashions many instances its measurement. The thesis behind it — for safety, a small mannequin skilled for the precise activity beats a large general-purpose one, and inference stays in-house so your supply by no means leaves the machine. Antares-350M and Antares-1B are publicly launched. Inside a month, they’ve reached 29,800 downloads on Hugging Face.
Antares Quickstart (cookbook).The sensible companion: a CLI that turns the mannequin into one thing you possibly can run in ~quarter-hour. It profiles a repo, checks for related weaknesses, and produces experiences (together with SARIF for GitHub Code Scanning), with privacy-preserving defaults and CI gating. That is the bridge from “spectacular benchmark” to “factor you gate each push with.”
Antares within the Wild (weblog, Jul 28).Every week after launch, the crew ran the unmodified fashions on two benchmarks constructed by different individuals: a Berkeley cyber-capability suite and a group rebuild of Devin’s safety eval. The discovering: localization holds up off-distribution, matching hosted brokers that price orders of magnitude extra per run. The potential generalizes, not simply suits one benchmark.
New Reasoning Benchmarks in FAITH (weblog, Jul 17).Two additions to our analysis hub, CTI-Reasoning and CWE-Prediction, constructed to measure whether or not a mannequin can purpose about safety slightly than recite memorized details: utilizing threat-intelligence evaluation duties and vulnerability information too latest to have been skilled on. Each are intentionally laborious and unsaturated, leaving headroom even for frontier fashions.
Black Hat USA 2026 (Aug 4–6).The Basis AI crew introduced Antares and our analysis on totally automated immediate optimization to Las Vegas, presenting at Cisco sales space classes and demonstrating the work on the AI Kiosk. We gathered beneficial suggestions from Antares customers whereas introducing the mannequin household to extra safety researchers, penetration testers, and safety engineers. These conversations bolstered Antares’s core concept: compact, task-specific fashions can ship sensible safety worth whereas protecting delicate code native. With agent analysis and enchancment rising as main themes throughout the convention, the occasion additionally underscored the potential of our open-source immediate optimization venture to assist autonomous brokers repeatedly evolve. The curiosity in immediate optimization additional demonstrated how open-sourcing this functionality can provide the group a sensible technique to construct brokers that be taught from analysis outcomes and enhance with much less guide tuning.
