Finding a vulnerable site in a large repository often takes longer than an analysis of the error itself, and Cisco tried to shorten this stage with the Antares compact AI model family. The company opened The weights of the Antares-350M and Antares-1B, which define files with a likely defect to describe a known vulnerability.
The models are not designed for automatic correction of the code, but for primary selection. Antares gets a description of the problem, searches for suitable templates, view possible files, takes into account the found connections and changes the direction if the chosen path does not give the result. At the output, the system forms a ranked list of files and saves the sequence of commands that led to the withdrawal.
The compact size allows you to run Antares locally or within the corporate infrastructure. The source code in such a scenario does not need to be transferred to the cloud service, and computing costs remain lower than when working with large universal models. Cisco expects that the models will be able to accelerate the analysis of security bulletins, checking individual error classes and early sorting of finds in development conveyors.
For evaluation, the developers created the Vulnerability Localization Benchmark of 500 tasks. Models were required to navigate in unfamiliar repositories and find files related to specific categories of weak CWE locations. According to Cisco, Antares surpassed a number of larger open and closed models, spending less time and computing resources. The results so far reflect the work on a special test set, and do not replace verification in real projects.
Cisco emphasizes that the Antares only narrows the search area. Teams still need dependency analysis, screening, dynamic checks, container control and infrastructure, threat modelling, and hand-held withdrawal checks. The Outlets-350M and Antares-1B are published on Hugging Face, and the company plans to introduce a larger Antares-3B company later.
The models are not designed for automatic correction of the code, but for primary selection. Antares gets a description of the problem, searches for suitable templates, view possible files, takes into account the found connections and changes the direction if the chosen path does not give the result. At the output, the system forms a ranked list of files and saves the sequence of commands that led to the withdrawal.
The compact size allows you to run Antares locally or within the corporate infrastructure. The source code in such a scenario does not need to be transferred to the cloud service, and computing costs remain lower than when working with large universal models. Cisco expects that the models will be able to accelerate the analysis of security bulletins, checking individual error classes and early sorting of finds in development conveyors.
For evaluation, the developers created the Vulnerability Localization Benchmark of 500 tasks. Models were required to navigate in unfamiliar repositories and find files related to specific categories of weak CWE locations. According to Cisco, Antares surpassed a number of larger open and closed models, spending less time and computing resources. The results so far reflect the work on a special test set, and do not replace verification in real projects.
Cisco emphasizes that the Antares only narrows the search area. Teams still need dependency analysis, screening, dynamic checks, container control and infrastructure, threat modelling, and hand-held withdrawal checks. The Outlets-350M and Antares-1B are published on Hugging Face, and the company plans to introduce a larger Antares-3B company later.
