Introducing Antares: Highly Efficient Open-Weight AI Models for Vulnerability Localization

Cisco has introduced Antares, a family of small language models (SLMs) purpose-built for vulnerability localization, the task of pinpointing where known flaws exist within a codebase. The first two releases, Antares-350M and Antares-1B, are open-weight models available now on Hugging Face, with benchmark testing showing they outperform many larger closed and open-weight models on this security task at a fraction of the cost. A third model, Antares-3B, is described as coming soon. Because the models are compact, they can run locally, which avoids sending sensitive codebases to external cloud services.

Why a small model for a costly security workflow

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 relevant evidence rarely sits 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.

Organizations also need security tooling that is practical to deploy. Compact models reduce inference costs, support local or on-premises operations, and help teams keep sensitive source code inside their own environment. Cisco positions Antares as a way to extend AI-assisted security to universities, public-sector institutions, and smaller security teams that have lacked the resources to run token-heavy models on critical tasks.

Software has always shipped with vulnerabilities, and large bodies of legacy code still carry latent flaws. AI now pressures the problem from both directions: agents write more of the code, and they are growing capable of exploiting it. That makes cheap, continuous verification a control that needs to sit close to the development loop, checking code before it lands. Small models are well suited here because they run locally, so proprietary code never leaves the machine, and they are fast enough to gate an agent’s output in real time.

Antares-1B is cited as an example: it recognizes vulnerabilities in code it is shown across several weakness classes and languages. A provided command-line interface (CLI) packages the model’s capabilities into a practical workflow for scanning codebases, inspecting findings, and integrating into automated pipelines.

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.”

Amin Saberi, Professor of Management Science and Engineering and Director of the Language, Data, and Reasoning Lab at Stanford University, added: “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.”

What makes Antares different

Traditional static analysis tools are valuable but often rule-heavy and can produce results that require substantial triage. General-purpose coding models can reason about code, but they are not necessarily optimized for security investigation, terminal navigation, or structured vulnerability localization. Antares is aimed at that middle ground.

The team says Antares was inspired by earlier 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 rather than from raw model scale alone.

Antares uses an iterative search pattern that mirrors 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. Cisco frames the goal as making the first stages of source-code vulnerability triage faster, more repeatable, and easier to review, rather than replacing expert judgment.

Workflows Antares can support include:

• 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 continuous integration and continuous deployment (CI/CD) triage workflows where likely vulnerable files need early review.
• Enabling local security analysis in environments with strict privacy or compliance requirements.

Antares outputs a ranked list of source files likely to contain a relevant vulnerability, along with the terminal exploration trace that led to that result. Security teams can use that output to prioritize analyst review during advisory triage, focused CWE investigations, or CI/CD security checks. Cisco is careful to note that Antares is not a replacement for the broader application security toolchain: dependency and software composition analysis, secret scanning, dynamic testing, infrastructure and container checks, threat modeling, remediation workflows, and expert review still all apply.

Why a new benchmark was needed

Building a model for vulnerable code localization raised an immediate measurement problem. General coding benchmarks were insufficient because Antares is not designed for general issue resolution, patch generation, or broad code search; it targets security-specific localization, identifying source files most likely to contain a vulnerability given limited security context.

The closest adjacent work is CodeScout. According to the CodeScout paper, posted on arXiv as “CodeScout: An Effective Recipe for Reinforcement Learning of Code Search Agents” (arXiv:2603.17829, submitted 18 March 2026, authored by 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), the project shows that, with an effective reinforcement learning recipe, a coding agent equipped with nothing more than a standard Unix terminal can be trained to achieve strong results. The study reports that CodeScout models consistently match or beat base and post-trained LLMs that are 2 to 18 times larger, and sometimes approach the performance of closed models like Claude Sonnet even when those models use specialized scaffolds, on the SWE-Bench Verified, Pro, and Lite benchmarks. CodeScout evaluates terminal-based code-search agents on software-engineering benchmarks, not on whether an agent can localize vulnerable files from CWE-style security descriptions, advisories, or vulnerability classes.

That gap led Cisco to build 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.

Why open-weight

Cisco is releasing Antares as open-weight because, the company says, the security community needs more accessible building blocks for practical, repository-level defense. Vulnerability triage is expensive in expertise, time, infrastructure, and access to sensitive code. Those costs are felt most sharply by organizations that already operate with constrained security resources, including universities, research institutions, nonprofits, and public-sector teams that still maintain important software.

How Antares fits into Cisco’s broader security AI effort

Antares is part of a wider Cisco push to advance how security organizations use AI responsibly. 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. With Antares, Cisco adds compact security models and a benchmark focused on repository-level vulnerability localization.

Cisco frames these efforts around a common belief: AI in security has to move beyond impressive one-off demos and toward systems that practitioners can evaluate, govern, and improve. Foundry Security Spec helps define the harness around agentic security work. CodeGuard helps raise the floor for prevention by capturing secure coding practices in a reusable rules corpus. Antares helps with the expensive localization step, narrowing vulnerability intelligence into source files that human reviewers can inspect.

Antares models are available on the Hugging Face page for community exploration, alongside the model card. Additional methodology details are in the technical paper, and Cisco points interested readers to its Foundation AI team for follow-up.

FAQ

What is Antares?

Antares is a family of small language models introduced by Cisco on July 21, 2026, purpose-built for vulnerability localization. The first two releases, Antares-350M and Antares-1B, are open-weight models available on Hugging Face, with Antares-3B described as coming soon.

Why does Cisco say Antares matters for security teams?

Cisco says Antares is designed to make source-code vulnerability triage faster and more repeatable at a fraction of the cost of larger models, and to run locally so sensitive code does not have to leave an organization’s environment.

What is the Vulnerability Localization Benchmark?

It is a 500-task benchmark introduced alongside Antares that requires models to navigate unfamiliar codebases and recognize vulnerability patterns tied to specific CWE categories, filling a gap that general coding benchmarks and adjacent work like CodeScout do not directly cover.

This article summarizes reporting from blogs.cisco.com. See our editorial disclaimer for how our articles are produced.