Cisco officially released and open-sourced a new family of small artificial intelligence models on Tuesday, designed to help companies more efficiently search for security vulnerabilities in software code bases. This family of models, called Antares, is designed for investigating software code bases and can accurately identify where known vulnerabilities are located in the code.
In a new benchmark that measures a model's ability to accurately identify vulnerable files in real-world code bases, Antares performed on par with larger models including OpenAI's GPT-5.5 and Z.ai's GLM-5.2, said Amin Karbasi, vice president and chief AI scientist at Cisco.
At the practical application level, running large AI models to fully scan an organization's large code base is often costly, and defenders need to repeat scans as the software is continuously updated. In contrast, Antares has demonstrated significant cost and efficiency advantages. In Cisco's tests, Antares scanned 500 code bases in only about 15 minutes at a cost of less than $1, while GPT-5.5 took about 5 hours and cost more than $10 billion. Because these smaller models can be run locally, organizations are able to keep sensitive source code within their own secure environment without sending it to a third-party AI service provider. DJ Sampass, senior vice president and general manager of Cisco AI Software and Platform, pointed out that enterprises need more practical tools in practical applications rather than overly redundant large-scale solutions.
In terms of training logic, Cisco did not train Antares to be an ordinary coding assistant, but to make it more like a professional security investigator. Rather than a general-purpose language model that simply answers questions about code, this model can autonomously learn how to search within a code base, inspecting files, pivoting when specific clues are blocked, and ultimately targeting the files most likely to contain vulnerabilities. Calbasi described this mechanism as like searching for security stitches in a vast ocean of code bases.
In terms of model security and release strategies, Cisco has cooperated with relevant U.S. government agencies on model security and release issues. To prevent malicious attackers and cybercriminals from conspiring to obtain relevant tools, Cisco will conduct rigorous review of users who download the models. Currently, Antares-350M and Antares-1B are available to verified cyber defenders on the Hugging Face platform on Tuesday. However, the more powerful 3 billion parameter version Antares-3B is not currently available to the public, and Cisco plans to integrate it directly into its own security product line.
Cisco's open source initiative is not an isolated case. Just last week, Capital One also open-sourced VulnHunter, an agent-based AI security tool that adopts an attacker's mindset and reviews source code to find potential hacking paths. Sampath revealed that Cisco is currently exploring the establishment of an industry alliance to further expand its work in the field of open source AI security tools, and this is just a starting point.