Cisco Debuts Antares AI Models to Streamline Local Vulnerability Detection

Cisco has introduced a new pair of specialized artificial intelligence models, known as Antares-350M and Antares-1B, designed to assist security professionals in locating specific weaknesses within extensive software codebases. Launched on July 21, 2026, these open-weight models function as terminal-based search agents that help narrow down the search for vulnerabilities. By running locally, the tools allow enterprises to scan sensitive or proprietary source code without the security risks associated with transmitting data to external cloud-based AI providers.

Rather than acting as comprehensive security scanners, the Antares models serve as initial investigative tools. When provided with a Common Weakness Enumeration (CWE) identifier and a description, the system utilizes standard commands like grep and find to examine a read-only version of a code repository. The software then generates a list of files that likely contain the flaw, outputting the data in human-readable, JSON, or SARIF formats. This workflow is intended to provide leads for human analysts who must still verify the flaws, as the models do not independently confirm vulnerabilities, determine their severity, or produce code fixes.

Hardware requirements for the models vary based on the deployment scale. The smaller Antares-350M version features a context window of 32,000 tokens, making it suitable for limited environments. In contrast, the Antares-1B model supports a 128,000-token window and is optimized for use on a single GPU. Cisco notes that these models have not undergone separate safety alignment for conversational tasks, advising users to secure the analysis environment and strictly control access to the model files and dependencies.

Evaluation data released by Cisco highlights both the potential and the current restrictions of this technology. Testing on the VLoc Bench—a dataset spanning 500 tasks across various software ecosystems—showed the Antares-1B model achieving a File F1 score of 0.209. While an unreleased 3B version of the model performed nearly as well as the top-tier GPT-5.5 configuration, overall benchmark scores remained below 0.23, with many complex tasks left unsolved by all tested AI systems. Because performance tends to drop as repositories increase in size or complexity, Cisco recommends that organizations treat Antares as an advisory tool rather than a replacement for traditional static analysis or human-led remediation.

---

Read original at TechRepublic AI.