Cisco releases Antares open-weight models for vulnerability localization
Cisco released Antares-350M and Antares-1B on July 21 as open-weight models for locating files associated with known vulnerability classes in software repositories. Cisco says both are available to the broader community on Hugging Face and can run locally; independent reporting says the larger Antares-3B is expected later. On Cisco's 500-task benchmark, 190 tasks remained unsolved by every evaluated system, underscoring that the models are triage aids, not replacements for application-security testing.
Cisco released Antares-350M and Antares-1B on July 21 as open-weight language models built for repository-level vulnerability localization. The company says both models are available to the broader community through Hugging Face and are small enough to run inside an organization's own environment. A larger Antares-3B model is expected later, according to Help Net Security and Cisco's launch materials.
What the models do
Antares is designed to identify which files are most likely to contain a known class of vulnerability. Given a repository and a Common Weakness Enumeration description, the model explores the codebase through a read-only terminal, inspects candidate files, revises its search path, and returns a ranked list of likely vulnerable files with an exploration trace.
That makes Antares a triage tool, not an automatic patcher or a general vulnerability scanner. Cisco's launch post says established controls—including dependency analysis, secret scanning, dynamic testing, threat modeling, and expert review—still belong in the workflow. Local inference may be useful where proprietary source code cannot be sent to an external model provider.
What Cisco's benchmark shows
Cisco evaluated the models on its Vulnerability Localization Benchmark, a company-created set of 500 tasks drawn from 290 repositories and 147 CWE categories. Each task pairs a repository snapshot containing a known flaw with file-level labels derived from the patch. A second phase tests whether a model avoids false alarms after the flaw has been fixed.
The accompanying technical report gives Antares-3B a File F1 score of 0.223, close to the 0.229 reported for GPT-5.5 under the same setup. Cisco's benchmark site also says 190 of the 500 tasks remained unsolved by every evaluated system. Those figures are useful for comparing systems within Cisco's protocol, but they have not been independently replicated, and the closest-performing Antares model is not one of the two weights released on July 21.
Practical use and limits
The release is evidence that task-specific training can make compact models competitive in a narrow security workflow. It does not show that a clean Antares result means a repository is safe, nor does file localization establish exploitability or severity.
Teams evaluating the models should measure recall on their own languages and repository sizes, review the exploration trace, and require reproducible evidence before opening or closing a remediation ticket. The benchmark's 38% all-model failure rate is the clearest guardrail: Antares may reduce the search space, but human validation and the rest of the application-security toolchain remain necessary.
Key Points
- 1Cisco released Antares-350M and Antares-1B broadly on Hugging Face; the larger Antares-3B is expected later rather than part of the current release.
- 2Cisco's 500-task benchmark spans 290 repositories and 147 CWE categories, but 190 tasks remained unsolved by every evaluated system.
- 3The models rank likely vulnerable files and can run locally, but Cisco presents them as triage aids that complement established security testing and expert review.
Scoring Rationale
Cisco released two usable compact security models with a public technical report and benchmark, giving application-security teams a concrete local-inference option. Impact is tempered because the benchmark is company-produced, the strongest Antares-3B model is forthcoming rather than currently released, and 38% of tasks defeated every evaluated system.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
