ME AI message, July 22 (UTC+8): Cisco has announced the launch of the Antares series of small language models, specifically designed for vulnerability detection. The official 350M and 1B parameter versions are now available on Hugging Face (click here to access), with a 3B parameter version to follow.
According to the introduction, Antares is not an AI model capable of autonomously discovering unknown vulnerabilities; its design purpose is to assist security personnel in vulnerability localization. After receiving the CWE vulnerability type and related description, the model uses terminal commands to search the codebase and generates an inspection report to help security personnel improve the efficiency of manual reviews.
To evaluate the Antares model’s performance, Cisco conducted a Vulnerability Localization Benchmark covering 290 real codebases, 147 CWE vulnerability types, and 500 test tasks. According to Cisco’s published technical report, Antares-1B achieved a 22.4% recall rate for identifying vulnerable files, ranking highest among all tested models, slightly above GPT-5.5’s 22.1%.
Cisco stated that due to its smaller scale, the Antares model can be directly deployed and run in local environments and enterprise networks, offering lightweight and privacy-focused characteristics. However, its functionality is limited and it can only serve as an auxiliary tool; enterprises still need to combine it with component analysis, software bill of materials (SBOM), sensitive data scanning, dynamic security testing, threat modeling, vulnerability remediation, and manual security reviews to establish a comprehensive software security protection system.(Source: X platform)