Cisco Antares Open-Weight Models Reduce Vulnerability Localization Costs

Post Views: 10

Cisco’s open-weight Antares models enhance vulnerability localization efficiency

Antares is engineered to operate on local or on-premises infrastructure

The models prioritize on-premises deployment. This approach reduces costs per task and aligns with the needs of institutions such as universities, public-sector entities, research organizations, and smaller security teams. By leveraging compact architectures, Antares provides a scalable alternative for vulnerability identification.

Model functionality mirrors human investigative processes

Each Antares variant processes repositories similarly to how a security analyst would. Starting with a vulnerability description, the models search for matching code patterns, evaluate candidate files, incorporate new evidence, backtrack when necessary, and refine their focus on critical files. The output includes a prioritized list of source files alongside the diagnostic path that led to their identification.

Antares integrates into existing security workflows

The tool addresses a single stage of the application security lifecycle, complementing established practices such as dependency analysis, secret scanning, dynamic testing, threat modeling, and manual reviews. Cisco developed the Vulnerability Localization Benchmark, a dataset of 500 entries, to evaluate models based on CWE-style security advisories rather than general coding tasks.

Performance metrics and cost efficiency

On the vulnerability localization benchmark, Antares models achieve high F1 scores, competing with larger systems. Antares-3B, the upcoming variant, approaches the performance of leading closed-source models like GPT-5.5. Antares-1B outperforms models such as GLM-5.2 and Gemini 3 Pro while operating at a fraction of their size. Cost analysis reveals significant savings: GPT-5.5 incurs approximately $141 per evaluation, whereas Antares reduces this by 172 times.

Impact on accessibility and security practices

The cost advantages of Antares address disparities in AI-driven security adoption, according to Amin Saberi, a Stanford University professor. He emphasized that advanced AI tools should not remain exclusive to organizations with substantial budgets. Antares enables near-frontier accuracy in secure code analysis at a fraction of the cost, allowing continuous security scanning for all teams.

Antares expands Cisco’s security initiatives

The release aligns with Cisco’s prior projects targeting similar challenges, including Foundry Security Spec, a framework for agent-based security evaluation, and CodeGuard, a set of rules for secure-by-default coding practices. The 350M and 1B variants are currently available on Hugging Face, with the 3B model forthcoming.

According to Amin Saberi, a Stanford University professor, “Advanced AI tools should not remain exclusive to organizations with substantial budgets.”