Cisco Launches Low-Cost AI Models for Source Code Security
Cisco Foundation AI released Antares, a small language model (SLM) family consisting of Antares-350M and Antares-1B, specifically optimized for identifying known vulnerabilities in codebases. The models utilize a human-like investigative search pattern to navigate repositories, aiming to reduce false positives compared to general language models while maintaining lower costs than closed frontier LLMs. Antares is positioned as a cost-effective, data-sovereign solution for security teams, particul
Analysis
TL;DR
- Cisco Foundation AI released Antares, a small language model (SLM) family consisting of Antares-350M and Antares-1B, specifically optimized for identifying known vulnerabilities in codebases.
- The models utilize a human-like investigative search pattern to navigate repositories, aiming to reduce false positives compared to general language models while maintaining lower costs than closed frontier LLMs.
- Antares is positioned as a cost-effective, data-sovereign solution for security teams, particularly benefiting organizations with limited budgets such as universities and public sectors.
- Performance was validated using the new Vulnerability Localization Benchmark, showing Antares to be significantly faster and cheaper than competitors like GPT-5.5 and GLM-5.2.
Why It Matters
This release addresses a critical gap in AI-driven cybersecurity by offering a specialized, open-weight alternative that balances cost, privacy, and accuracy. For AI practitioners and security engineers, it demonstrates the viability of Small Language Models (SLMs) for complex, domain-specific tasks like code analysis, challenging the assumption that only massive frontier models can handle sophisticated security reasoning.
Technical Details
- Model Architecture: The Antares family includes two open-weight SLMs, Antares-350M and Antares-1B, available on Hugging Face.
- Operational Mechanism: Unlike generic text generation, Antares employs a search pattern mimicking human investigation: starting from a vulnerability description, searching for code patterns, reading candidate files, incorporating evidence, and narrowing down to high-probability files.
- Benchmarking: Cisco introduced the Vulnerability Localization Benchmark, a 500-entry dataset requiring models to navigate unfamiliar codebases and recognize CWE-associated vulnerability patterns.
- Performance Metrics: In benchmark tests, Antares outperformed OpenAI’s GPT-5.5 and Z.ai’s GLM-5.2 in speed and cost-efficiency, being 172x cheaper than the closed LLM and 15.2x cheaper than the open-weight GLM.
Industry Insight
- Democratization of Secure AI: Organizations with strict data sovereignty requirements or limited budgets can now deploy advanced AI security tools locally without relying on expensive, external API calls to frontier models.
- Shift Towards Specialized SLMs: The success of Antares suggests a future trend where highly specialized Small Language Models will replace general-purpose LLMs for specific enterprise tasks, offering better ROI and reduced hallucination rates.
- Standardization of Security Benchmarks: The introduction of the Vulnerability Localization Benchmark provides a standardized metric for evaluating AI performance in code security, encouraging more rigorous and comparable testing across the industry.
Disclaimer: The above content is generated by AI and is for reference only.