Cisco’s Foundation AI team has released Antares, a series of compact, open-weight artificial intelligence models intended for agentic vulnerability localization. The model family includes versions with 350 million, 1 billion, and 3 billion parameters. These systems are designed to pinpoint specific files containing security weaknesses within a codebase using only a Common Weakness Enumeration description and read-only repository access.
The 3 billion parameter version is reported to rival larger frontier systems like GPT-5.5 while operating locally at a significantly lower cost. It also outperforms open-source models that are substantially larger. The underlying approach suggests that smaller models trained specifically for security tasks can be more effective than general-purpose giants, allowing source code to remain on local machines.
The 350 million and 1 billion parameter models are publicly available and have accumulated nearly 30,000 downloads on Hugging Face within one month of release.
To facilitate practical use, the team released an Antares Quickstart cookbook. This includes a command-line interface that allows users to profile repositories, check for weaknesses, and generate reports in formats such as SARIF for GitHub Code Scanning. The tool is designed to integrate into continuous integration pipelines with privacy-preserving defaults.
Recent testing on external benchmarks, including a Berkeley cyber-capability suite and a community-built security evaluation, indicated that the models maintain performance outside their training distribution. These results matched the capabilities of hosted agents that incur much higher per-run costs. Additionally, the team introduced two new reasoning benchmarks, CTI-Reasoning and CWE-Prediction, to the FAITH evaluation hub.
These tests measure a model's ability to reason about security using recent data rather than relying on memorized facts.