Cisco has released Antares, a family of security-focused small language models designed to pinpoint where known vulnerabilities exist inside a codebase. The first two models, Antares-350M and Antares-1B, are open-weight and available on Hugging Face, with a third model, Antares-3B, announced as coming soon. According to Cisco, the models outperform larger open- and closed-weight alternatives on a new vulnerability localization benchmark while running at a fraction of the cost and small enough to operate locally, which keeps proprietary source code off the cloud.<\/p>\n
Software security depends on connecting external knowledge, including public databases, advisories, and Common Weakness Enumeration (CWE) categories, to internal code. That work is hard because repositories are large, security signals are noisy, and the relevant evidence is rarely in one obvious place. Analysts often have to search unfamiliar code, follow naming conventions, inspect call paths, compare candidate files, and decide whether a weakness is actually present.<\/p>\n
General-purpose coding models can reason about code, but they are not necessarily optimized for security investigation, terminal navigation, or structured vulnerability localization. Traditional static analysis tools are rule-heavy and produce results that can require substantial triage. Antares is designed for the middle ground between these two approaches.<\/p>\n
The team also pointed to a second motivation: the economics of security. Compact models reduce inference costs, can run on premises, and let teams keep sensitive source code inside their own environment. That, the authors write, can make AI-assisted security viable for universities, public-sector institutions, and smaller security teams that previously could not afford token-heavy frontier models for the task.<\/p>\n
Antares follows an iterative search pattern modeled on how a human investigator works through a repository. Each model starts from a vulnerability description, searches for relevant code patterns, reads candidate files, incorporates new evidence, changes direction when a path is unhelpful, and narrows toward the files most likely to matter. Its output is a ranked list of source files likely to contain a relevant vulnerability, along with the terminal exploration trace that produced that ranking.<\/p>\n
The approach was inspired by prior research from the Cisco Foundation AI team showing that compact models can learn to search, reflect, revise strategy, and backtrack when a path is unproductive. The Cisco team wanted to test whether the same idea could be applied to an urgent real-world security problem: vulnerable code localization.<\/p>\n
The release of Antares-1B ships with a command-line interface that packages these capabilities into a workflow, making it possible to scan codebases, inspect findings, and integrate the model into automated pipelines.<\/p>\n
General coding benchmarks such as SWE-Bench Verified, Pro, and Lite evaluate whether an agent can find code relevant to a software issue or development task. The Cisco team argued that those benchmarks do not measure whether a model can localize vulnerable files from CWE-style security descriptions, advisories, or vulnerability classes.<\/p>\n
To fill that gap, Cisco introduced the Vulnerability Localization Benchmark, a 500-task benchmark that requires models to simultaneously navigate unfamiliar codebases efficiently and recognize vulnerability patterns tied to specific CWE categories.<\/p>\n
The closely related work, CodeScout, is documented in detail on arXiv. The paper, “CodeScout: An Effective Recipe for Reinforcement Learning of Code Search Agents” (arXiv:2603.17829, submitted 18 March 2026), shows that with an effective reinforcement learning recipe, a coding agent equipped only with a standard Unix terminal can be trained to achieve strong localization results. The CodeScout authors, listed on the paper as Lintang Sutawika, Aditya Bharat Soni, Bharath Sriraam R R, Apurva Gandhi, Taha Yassine, Sanidhya Vijayvargiya, Yuchen Li, Xuhui Zhou, Yilin Zhang, Leander Melroy Maben, and Graham Neubig, report that their models achieve superior or competitive performance compared with base and post-trained large language models 2 to 18 times larger, and sometimes approach the performance of closed models such as Claude Sonnet, even when those closed models use specialized scaffolds. The experiments were conducted on three benchmarks: SWE-Bench Verified, Pro, and Lite. The CodeScout paper focuses on techniques for re-purposing existing coding agent environments for code search, reward design, and reinforcement learning optimization, and the resulting model family, code, and data are released for the community.<\/p>\n
Cisco’s announcement builds on that line of work but tailors the evaluation to security-specific localization tasks: given limited security context, identifying source files most likely to contain a vulnerability.<\/p>\n
Cisco positions Antares as part of a coordinated effort to make AI useful for defenders, not just impressive in demos. The release is paired with two other Cisco projects:<\/p>\n
Taken together, Foundry Security Spec defines the harness around agentic security work, CodeGuard captures secure coding practices in a reusable rules corpus, and Antares handles the localization step, narrowing vulnerability intelligence into source files that human reviewers can inspect. The company also notes that Antares is not meant to replace the broader application security toolchain. Security teams still need dependency and software composition analysis, secret scanning, dynamic testing, infrastructure and container checks, threat modeling, remediation workflows, and expert review.<\/p>\n
Two academic researchers quoted in the announcement framed the efficiency story as the central contribution. Reza Shokri, Associate Professor of Computer Science at the National University of Singapore, said that the compact Antares-1B model could navigate a codebase and identify security issues that would otherwise require larger models or more manual analysis. Amin Saberi, Professor of Management Science and Engineering and Director of the Language, Data, and Reasoning Lab at Stanford University, said that “near-frontier accuracy on secure code reasoning at a fraction of the cost, fast enough to run on every commit” is what makes always-on security scanning possible for teams that cannot afford frontier-scale budgets.<\/p>\n
Cisco lists several practical workflows the model is intended to support:<\/p>\n
The goal, the authors write, is not to replace expert judgment, but to make the first stages of source-code vulnerability triage faster, more repeatable, and easier to review.<\/p>\n
Antares-350M and Antares-1B are available on Hugging Face with model cards describing capabilities and limits. A technical paper describes the benchmark and evaluation methodology. The work is led by Amin Karbasi, VP and Chief AI Scientist at Foundation AI, with collaborators including Supriti Vijay, Aman Priyanshu, Didier Chapoteau, Arthur Goldblatt, Kimia Majd, Fraser Burch, Jianliang He, Baturay Saglam, Takahiro Matsumoto, and Zhuoran Yang.<\/p>\n
Antares is a family of security-focused small language models from Cisco, designed to localize vulnerable source code inside repositories. The first two releases, Antares-350M and Antares-1B, are open-weight and available on Hugging Face, with Antares-3B announced as coming soon.<\/p>\n
Antares is purpose-built for security-specific localization. It starts from a vulnerability description or CWE category, iteratively searches the repository, reads candidate files, and outputs a ranked list of files most likely to contain the vulnerability, along with the terminal exploration trace that led to that ranking.<\/p>\n
Existing benchmarks such as SWE-Bench Verified, Pro, and Lite measure localization for general software issues, not for vulnerability-specific tasks. Cisco’s Vulnerability Localization Benchmark is a 500-task evaluation that requires models to navigate unfamiliar codebases and recognize vulnerability patterns tied to specific CWE categories, which the team argues is closer to the real workflow of security triage.<\/p>\n