Cisco Antares: Open-Weight AI Models for Vulnerability Localization

Free Instant SEO Audit

By NewSunSEO July 23, 2026

Cisco has introduced Antares, a family of open-weight small language models (SLMs) built specifically for vulnerability localization, the task of pinpointing where known vulnerabilities sit inside a codebase. On July 21, 2026 the company released Antares-350M and Antares-1B on Hugging Face, with Antares-3B announced as coming soon, and published a new 500-task Vulnerability Localization Benchmark to measure performance on the task.

Why a small model for a costly workflow?

Software security depends on linking external knowledge such as public vulnerability databases, advisories, and Common Weakness Enumerations (CWEs) to internal code. Repositories are large, security signals are noisy, and the relevant evidence is rarely in one obvious place. Analysts typically have to search unfamiliar code, follow naming conventions, inspect call paths, compare candidate files, and decide whether a weakness is actually present.

Compact models lower inference cost, support local or on-premises deployment, and keep proprietary source code inside an organization’s own environment. The release targets universities, public-sector institutions, and smaller security teams that have struggled to afford token-heavy frontier models for routine checks. Amin Karbasi, VP and Chief AI Scientist for Foundation AI at Cisco, framed the timing in his blog post: as AI agents write more code and grow more capable of exploiting it, cheap continuous verification has to sit close to the development loop, checking code before it lands.

What Antares does

Antares is not a general code-generation model. It accepts a vulnerability description and outputs a ranked list of source files likely to contain a relevant vulnerability, along with the terminal exploration trace that led to that ranking. The model follows an iterative search pattern: starting from a vulnerability description, it 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 provided command-line interface packages those capabilities into a workflow that can scan codebases, inspect findings, and integrate into automated pipelines. Cisco lists intended use cases that include locating files related to a CWE category in a repository snapshot, triaging advisory-driven investigations, augmenting static analysis with model-driven repository exploration, supporting CI/CD triage where likely vulnerable files need early review, and enabling local analysis in environments with strict privacy or compliance requirements.

How it compares on the new benchmark

Cisco’s Vulnerability Localization Benchmark is a 500-task evaluation that requires models to navigate unfamiliar codebases while recognizing patterns tied to specific CWE categories. According to the blog, the Antares family outperforms a dozen larger models tested in the benchmark, at a substantially lower estimated cost and runtime. Cisco attributes that result in part to prior work by its Foundation AI team showing that compact models can learn to search, reflect, revise strategy, and backtrack, meaning useful retrieval behavior does not require frontier scale.

External researchers quoted in the announcement highlighted the practical implications. Reza Shokri, Associate Professor of Computer Science at the National University of Singapore, said that Antares-1B is impressive because it can 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, argued that near-frontier accuracy on secure-code reasoning at a fraction of the cost makes always-on scanning feasible for every team, calling it a necessity rather than a luxury.

How the benchmark differs from prior work

Cisco notes that general coding benchmarks measure whether an agent can find code relevant to a development task, not whether it can localize vulnerable files from CWE-style descriptions, advisories, or vulnerability classes. The closest adjacent work, according to Cisco, is CodeScout, a separate research effort described in an arXiv paper titled “CodeScout: An Effective Recipe for Reinforcement Learning of Code Search Agents,” submitted March 18, 2026 by Lintang Sutawika and ten co-authors.

The CodeScout paper, published on arXiv as 2603.17829 in the Computer Science software engineering category, focused on training code-search agents with reinforcement learning using only a standard Unix terminal. CodeScout’s experiments covered three benchmarks, SWE-Bench Verified, Pro, and Lite, and reported that the resulting models achieved superior or competitive performance against base and post-trained LLMs that were 2-18x larger, occasionally approaching closed models such as Claude Sonnet. The CodeScout authors released their model family along with code and data. Cisco’s blog points to CodeScout as evidence that code localization can be trained and measured with a standard Unix terminal, while arguing that security-specific localization needed its own benchmark, which the Vulnerability Localization Benchmark is designed to provide.

Why open-weight

Cisco is releasing Antares as open-weight rather than closed because, in the company’s framing, the security community needs more accessible building blocks for practical, repository-level defense. Vulnerability triage requires expertise, time, infrastructure, and access to sensitive code, costs that fall hardest on universities, research institutions, nonprofits, and public-sector teams that maintain important software. The Hugging Face listing makes the weights and a model card available for community review.

Where Antares fits in the broader toolchain

Cisco is explicit 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. The blog positions Antares alongside two earlier Cisco efforts: Foundry Security Spec, a model-agnostic blueprint for building agentic security evaluation systems with clear roles, guardrails, and reviewable outputs, and CodeGuard, a corpus of secure-by-default rules and skills intended to guide AI coding agents toward safer code. Foundry Security Spec is described as defining the framework around agentic security work, CodeGuard as raising the baseline for prevention, and Antares as narrowing vulnerability intelligence into source files that human reviewers can inspect.

The stated goal is a system in which every security practitioner, regardless of on-premises or resource constraints, can incorporate AI into everyday operations. Readers can explore the Antares models and model card on Hugging Face or read the technical paper for benchmark and evaluation methodology details. The blog lists Cisco Foundation AI as the contact point for further inquiries.

FAQ

What are the Antares models and who released them?

Antares is a family of open-weight small language models for vulnerability localization released by Cisco on July 21, 2026. The initial release includes Antares-350M and Antares-1B on Hugging Face, with Antares-3B announced as coming soon.

What is the Vulnerability Localization Benchmark?

It is a 500-task benchmark published by Cisco alongside the Antares release. It requires models to navigate unfamiliar codebases while recognizing vulnerability patterns tied to specific CWE categories, measuring a security-specific localization skill that Cisco says general coding benchmarks do not cover.

Why did Cisco release Antares as open-weight?

Cisco’s stated reason is that vulnerability triage is expensive and requires access to sensitive code, costs that fall hardest on universities, research institutions, nonprofits, and public-sector teams. Releasing the weights on Hugging Face lets those groups run the model locally without sending code to the cloud.