Cisco has released Antares, a family of small language models designed specifically for one of the most expensive and time-consuming tasks in software security: pinpointing where a known vulnerability lives inside a codebase. The two initial models, Antares-350M and Antares-1B, are published as open-weight releases on Hugging Face, with a larger Antares-3B announced as coming soon. Cisco’s benchmarking shows the family outperforming a dozen larger open-weight and closed models on its new Vulnerability Localization Benchmark while running at a fraction of the estimated cost and runtime.<\/p>\n
Software security depends on connecting external vulnerability knowledge, including public databases, advisories, and Common Weakness Enumerations (CWEs), 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 typically search through unfamiliar code, follow naming conventions, inspect call paths, compare candidate files, and decide whether a weakness is actually present.<\/p>\n
Antares is designed for that middle ground between traditional static analysis, which can be rule-heavy and produce results that need substantial triage, and general-purpose coding models, which may not be optimized for security investigation or structured vulnerability localization. The model family is meant to accelerate the first stages of source-code vulnerability triage rather than replace expert judgment.<\/p>\n
Antares follows an iterative search pattern that resembles 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. The output is a ranked list of source files likely to contain a relevant vulnerability, along with the terminal exploration trace that led to that result.<\/p>\n
In practice, the models can support workflows such as locating files related to a CWE category in a repository snapshot, triaging advisory-driven security investigations, augmenting static analysis with model-driven repository exploration, supporting CI\/CD triage where likely vulnerable files need early review, and enabling local security analysis in environments with strict privacy or compliance requirements.<\/p>\n
The approach was inspired by prior Cisco Foundation AI research showing that compact models can learn to search, reflect, revise their strategy, and backtrack when a path is unproductive. In other words, useful retrieval behavior can come from learned search strategies, not only from model scale. A provided command-line interface packages the model’s capabilities into a workflow that scans codebases, inspects findings, and integrates with automated pipelines.<\/p>\n
General coding benchmarks do not measure what Antares is built for. The closest adjacent work is CodeScout, a research project that evaluates terminal-based code-search agents on software-engineering benchmarks such as SWE-Bench Verified, Pro, and Lite. The CodeScout paper, hosted on arXiv as “CodeScout: An Effective Recipe for Reinforcement Learning of Code Search Agents” by Lintang Sutawika and ten co-authors and submitted on 18 March 2026, demonstrates that a coding agent equipped with nothing more than a standard Unix terminal can match 2-18x larger base and post-trained LLMs on those benchmarks, and sometimes approach closed-model performance like Claude Sonnet, even when those models use specialized scaffolds. The benchmark focus in CodeScout is whether an agent can find code relevant to a software issue or development task, not whether it can localize vulnerable files from CWE-style security descriptions, security advisories, or vulnerability classes.<\/p>\n
To fill that gap, Cisco introduced the Vulnerability Localization Benchmark, a 500-task benchmark requiring models to simultaneously navigate unfamiliar codebases efficiently and recognize vulnerability patterns associated with specific CWE categories. According to Cisco’s chart, the Antares model family completes runs at substantially lower estimated cost and runtime than larger comparison models while still outperforming them on the task.<\/p>\n
Cisco is releasing Antares as open-weight because the security community needs more accessible building blocks for practical, repository-level defense. Vulnerability triage is expensive: it requires expertise, time, infrastructure, and access to sensitive code. Those costs hit hardest at universities, research institutions, nonprofits, and public-sector teams that still maintain important software.<\/p>\n
Compact models reduce inference costs, support local or on-premises operations, and help teams keep sensitive source code within their own environment. Antares-1B is highlighted as a concrete example: it recognizes vulnerabilities in code it is shown across several weakness classes and languages, and runs locally so proprietary code never leaves the machine. Cisco frames the release as a way to make always-on security scanning practical for every team, including those without frontier-scale budgets.<\/p>\n
Antares is one piece of a larger Cisco effort to advance how security organizations use AI. With Foundry Security Spec, Cisco shared a model-agnostic blueprint for building agentic security evaluation systems with clear roles, guardrails, and reviewable outputs. With CodeGuard, Cisco helped contribute secure-by-default rules and skills that can guide AI coding agents toward safer software development.<\/p>\n
Antares tackles the expensive localization step, narrowing vulnerability intelligence into source files that human reviewers can inspect. Together, the three efforts point at a more durable foundation for AI-assisted defense: open specifications, reusable security knowledge, compact deployable models, and benchmarks that make progress measurable.<\/p>\n
Reza Shokri, Associate Professor of Computer Science at the National University of Singapore, commented: “What I found compelling was seeing a model this small navigate a codebase and identify security issues that would otherwise require larger models or more manual analysis.”<\/p>\n
Amin Saberi, Professor of Management Science and Engineering and Director of the Language, Data, and Reasoning Lab at Stanford University, said: “Security can’t be a luxury good, yet advanced AI-based detection has largely belonged to organizations with frontier-scale budgets. Antares’s results change that equation: near-frontier accuracy on secure code reasoning at a fraction of the cost, fast enough to run on every commit. At a moment when attackers are already using AI, that kind of efficiency isn’t a nice-to-have; it’s what makes always-on security scanning possible for every team.”<\/p>\n
The Antares models and model card are available on Cisco’s Hugging Face page. A technical paper covers the benchmark and evaluation methodology. Inquiries can be directed to Cisco Foundation AI.<\/p>\n
Antares is a family of small language models from Cisco, purpose-built for vulnerability localization, the task of identifying which source files in a codebase are likely to contain a known vulnerability.<\/p>\n
Cisco has released Antares-350M and Antares-1B as open-weight models on Hugging Face, with a larger Antares-3B announced as coming soon.<\/p>\n
It is a 500-task benchmark introduced by Cisco that requires models to navigate unfamiliar codebases efficiently and recognize vulnerability patterns associated with specific CWE categories, going beyond what general-purpose coding benchmarks such as SWE-Bench Verified, Pro, and Lite measure.<\/p>\n