A new family of open-weight AI models helps developers locate vulnerable source files faster, enabling secure code analysis on local hardware at a fraction of the cost of larger language models.
A family of open-weight AI models has been introduced for vulnerability localisation. Developed by Cisco, the Antares series features models with 350 million and 1 billion parameters, a large 3-billion-parameter is also planned. Smaller models are hosted at Hugging Face and aim to cut down the time wasted on manually tracking down software vulnerabilities.
As opposed to existing solutions for searching within codebases, the models look into code in the way a human researcher would. From a description of the vulnerability, the algorithm searches for relevant code patterns, examines potential candidates, uses newly gathered information to continue narrowing down the search, backtracks when needed, and produces a list of top vulnerable files with an explanation of how they were identified.
Due to the relatively compact nature of these models, they can be run locally or on-premises hardware without the need for any cloud computing facilities. This allows organisations to keep sensitive source code locally while keeping the cost of inference low. These models are meant to help universities, research organisations, public sector organisations, as well as smaller security firms that do not have access to AI infrastructure.
For the assessment of performance, the developers have introduced the Vulnerability Localisation Benchmark (VLB), a dataset which consists of 500 realistic security cases where models have to explore unknown code bases and identify files associated with Common Weakness Enumeration (CWE) categories.
Benchmark numbers indicate that the forthcoming 3B model is comparable to the state-of-the-art AI models on the benchmark, whereas the 1B model significantly outperforms larger open-weight models like GLM-5.2 and Gemini 3 Pro even with many fewer parameters. Moreover, both small models incur notably less cost on evaluations, enabling developers to run continuous AI-powered vulnerability assessments.
The Antares family is designed to augment application security workflows rather than replace them. They are designed to work together with dependency analysis, software composition analysis, secret detection, dynamic testing, threat modeling, and expert review. Related resources, including Foundry Security Spec for agentic security evaluation and CodeGuard for secure-by-default coding, are also available for building AI-assisted software security pipelines.
