
Quick Answer
Antares is a family of compact, open-weight AI models that can run locally and identify source files likely to contain a known vulnerability. Cisco released Antares-350M and Antares-1B on Hugging Face on July 21, 2026, while describing Antares-3B as forthcoming; a later Cisco Model Foundry announcement lists Antares-3B in its model catalog.
In This Article
- What Business Teams Should Know
- What Happened
- Are You Affected?
- Safe Evaluation Checklist
- What to Do Now
- What the Evidence Shows
- How Antares Searches a Repository
- Why Open-Weight Models Matter for Repository Security
- What Changed Since the Release
- How Antares Fits the Security Stack
- Key Considerations
- FAQ
- Related Coverage
What Business Teams Should Know
- What Antares is: Cisco describes Antares as a family of security small language models built for repository-level vulnerability localization. Given limited security context, the models return a ranked list of source files and the terminal exploration trace used to find them.
- Can it run locally? Yes. Cisco says the compact models support local or on-premises operation, allowing teams to keep sensitive source code within their own environment.
- What does it not replace? Cisco says Antares does not replace dependency and software composition analysis, secret scanning, dynamic testing, infrastructure and container checks, threat modeling, remediation workflows, or expert review.
What Happened
Cisco introduced Antares on July 21, 2026 as a family of security small language models for locating known vulnerabilities inside code repositories. The company released Antares-350M and Antares-1B as open-weight models on Hugging Face, while Antares-3B was described as “coming soon” rather than released at that time.
Cisco’s August 3, 2026 technical report describes three model sizes, 350M, 1B, and 3B parameters. The report says Antares-3B approaches GPT-5.5 while outperforming open-weight models over 200x larger in size.
Are You Affected?
Antares applies to teams that maintain large software repositories, investigate advisories, perform CWE-specific searches, support static analysis or CI/CD triage, or need local analysis because of privacy and compliance requirements. Cisco specifically points to security teams, software developers, universities, public-sector organizations, nonprofits, and research institutions.
Antares-350M and Antares-1B are the models available on Hugging Face as of the release announcement. Antares-3B was described as forthcoming in July 2026. The October 26, 2026 Cisco Model Foundry announcement lists Antares-3B as the most capable model in the Antares family, purpose-built for agentic vulnerability localization.
Safe Evaluation Checklist
- Verify the model, documentation, and license against Cisco’s official model cards before use.
- Start with a controlled, non-production repository rather than a live production codebase.
- Compare ranked results with the team’s existing triage method.
- Record false positives, missed files, reviewer time, and operating cost.
- Require human confirmation before remediation.
- Retain complementary security controls.
What to Do Now
- Review the Antares-350M and Antares-1B model cards on Hugging Face and confirm that the licensing and documented usage conditions fit your environment.
- Compare the two available models with the repositories and weakness classes your team handles.
- Check whether the ranked files and terminal exploration trace help reviewers find the relevant implementation.
- Map the output to existing advisory investigations, focused CWE reviews, and CI/CD checks that need early human review.
What the Evidence Shows
The Vulnerability Localization Benchmark technical report, submitted to arXiv on September 14, 2026, contains 500 real-world vulnerabilities from 290 repositories, six package ecosystems, and 147 CWE categories. Each task pairs repository snapshots from immediately before and after a security fix. On the vulnerable snapshot, the agent receives a CWE description and read-only terminal access; on the patched snapshot, it must determine that the recorded vulnerability is no longer present.
The same report says 27 language models and four static-analysis tools were evaluated under a common agent interface. The strongest system achieved 0.229 File F1, and 38.4% of tasks received no correct localization from any evaluated model. The report also found that strong localization does not guarantee reliable behavior after remediation, because systems that identify vulnerable files can still report unsupported locations in patched repositories.
The Antares technical report says a full 500-task evaluation sweep takes approximately 15 minutes on one H100 GPU, with amortized evaluation time under two seconds and less than $0.002 per task.
According to a October 6, 2026 Cisco Foundation AI summary by Huaibo Zhao, AI Researcher, Antares-3B ranked second overall on VLoc Bench with 0.223 File F1, compared with 0.229 for the strongest system. The same summary notes that systems strongest at finding vulnerable files were not necessarily best at recognizing that those vulnerabilities had already been fixed.
How Antares Searches a Repository
Antares uses an iterative search process that Cisco compares with a human investigator. The model starts with a vulnerability description, searches for relevant code patterns, reads candidate files, incorporates new evidence, changes direction when a path is unhelpful, and narrows toward files likely to matter.
The output has two parts: a ranked list of source files likely to contain the vulnerability and the terminal exploration trace behind that ranking. Cisco says the goal is to make the first stages of source-code triage faster, more repeatable, and easier to review, while leaving final judgment with security personnel.
The provided CLI packages this workflow for scanning codebases, inspecting findings, and connecting the models to automated pipelines. Cisco positions Antares for advisory-driven investigations, CWE-specific repository searches, static-analysis support, CI/CD triage, and local analysis where cloud use is restricted.
Why Open-Weight Models Matter for Repository Security
Cisco released the models with open weights because repository-level triage requires security expertise, time, infrastructure, and access to sensitive code. In a July 24, 2026 Cisco partner article, the company says more than 20% of the most targeted vulnerabilities were over a decade old in the previous year, according to Cisco Talos.
Local deployment also gives organizations a way to run continuous checks without sending proprietary source code outside their environment. Cisco’s point is narrower than replacing frontier models: Antares focuses on finding where an already-known vulnerability lives in internal code.
What Changed Since the Release
Cisco’s October 26, 2026 Model Foundry announcement lists Antares-3B among the purpose-built models available to partners through Cisco Cloud Control, changing its status from the “coming soon” description in the July release post. The October 6, 2026 Foundation AI summary also reports Antares-3B at 0.223 File F1, second overall on VLoc Bench.
How Antares Fits the Security Stack
Cisco connects Antares with Foundry Security Spec, a model-agnostic blueprint for agentic security evaluation systems with defined roles, guardrails, and reviewable outputs. It also connects the models with CodeGuard, which contains secure-by-default rules and skills for AI coding agents.
Related research supports the general repository-search approach without establishing the same results for Antares. The March 18, 2026 CodeScout paper reports that models trained with an effective reinforcement learning recipe achieved superior or competitive performance against base and post-trained language models two to 18 times larger across SWE-Bench Verified, Pro, and Lite. The authors say the models sometimes approached performance from closed models such as Claude Sonnet, even when using specialized scaffolds.
Key Considerations
- The VLoc Bench results apply to a specific 500-task benchmark, and 38.4% of those tasks received no correct localization from any evaluated model.
- Antares-3B was described as forthcoming at the July 2026 release; the October 2026 Model Foundry announcement lists it in the partner model catalog.
- Performance and cost figures reported by the model authors depend on the benchmark, H100 hardware, and the evaluation setup.
- Strong file localization does not establish that a vulnerability was fixed, because evaluated systems sometimes reported unsupported locations in patched repositories.
- Cisco says Antares supports initial triage and does not replace dependency analysis, secret scanning, dynamic testing, infrastructure checks, threat modeling, remediation, or expert review.
FAQ
What Is Cisco Antares?
Antares is a family of compact security small language models for repository-level vulnerability localization. Cisco says the models identify source files likely to contain a known vulnerability and return the terminal exploration trace behind the ranking.
Can Antares Models Run Locally?
Yes. Cisco says the compact models support local and on-premises operation, allowing organizations to keep sensitive source code within their own environment.
Which Antares Models Are Available?
Cisco released Antares-350M and Antares-1B as open-weight models on Hugging Face on July 21, 2026. Antares-3B was described as coming soon in that announcement and was later listed in the October 26, 2026 Cisco Model Foundry model catalog.
Does Antares Replace Existing Security Tools?
No. Cisco says Antares does not replace dependency and software composition analysis, secret scanning, dynamic testing, infrastructure and container checks, threat modeling, remediation workflows, or expert review.
How Does Antares Perform on Vulnerability Localization?
According to the Antares technical report, Antares-3B approaches GPT-5.5 while outperforming open-weight models over 200x larger in size. A separate Cisco Foundation AI summary reports that Antares-3B ranked second overall on VLoc Bench at 0.223 File F1, while the strongest system reached 0.229.
Related Coverage
- Tech Firms Urge Washington Not to Restrict Open-Weight AI
- Poolside Laguna S 2.1 open-weight coding model released
Sources
See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.