Cisco Antares AI models hunt bugs faster than Gemini and GPT, cost under $1 to scan 500 repos

Cisco Antares Challenges Frontier Models with Compact Design

Cisco has released two open-weight AI models that specialize in vulnerability detection, challenging the notion that only expensive frontier systems can effectively hunt software bugs 1

1

. The Cisco Antares family includes Antares-350M and Antares-1B, both now available on Hugging Face to vetted users 1

1

. These small language models are purpose-built to pinpoint where known vulnerabilities exist within a codebase, addressing one of the most time-consuming and expensive problems in security 4

4

.

The networking giant's approach represents a quiet rebuke to the frontier-model arms race. "You really don't need a private jet to go to your corner store," DJ Sampath, Cisco's senior vice president and general manager of AI software and platform, told Axios . The models are designed to run locally, meaning proprietary code never leaves the organization's machines, unlike cloud-based systems that send code to external servers for processing 1

1

. This local deployment capability enables security analysis in environments with strict privacy or compliance requirements 1

1

.

Performance Metrics Show Speed and Cost Advantages

Cisco claims that Antares-1B outperforms Google Gemini 3 Pro and matches Z.ai's GLM-5.2 on its new benchmark for vulnerability localization 1

1

. The unreleased Antares-3B model reportedly surpasses both GLM-5.2 and OpenAI GPT-5.5 at finding vulnerabilities 1

1

. The real advantage appears in speed and cost-effective operation. Amin Karbasi, Cisco's vice president and chief AI scientist, revealed that Antares scans 500 software repositories in 15 minutes for less than $1, whereas frontier models take five hours and cost between $100 to $150 1

1

.

This efficiency matters for organizations that need to repeatedly search large codebases as software changes . Running large AI models across an organization's codebase can become expensive when defenders must rescan repositories continuously . The compact size of these AI-powered security tools makes AI-driven cybersecurity accessible to universities, public sector institutions, and smaller security teams that may have lacked resources to use token-intensive AI models 4

4

.

Trained as Investigator, Not Chatbot

What sets Cisco Antares apart is its fundamental architecture. "Antares is inherently not a chatbot. It is an investigator. It is a search engine," Karbasi explained 1

1

. The models were trained using learned search strategies that allow them to search, reflect, revise their approach, and backtrack when a path proves unproductive 4

4

. Rather than detecting a specific CVE or generating a patch, these models search a codebase using only a Common Weakness Enumeration description and return the files most likely to contain that class of vulnerability 2

2

.

"Its purpose is to reduce a large codebase to a focused set of files that a security professional or a downstream security workflow should investigate," Cisco AI researcher Supriti Vijay explained 2

2

. The goal centers on reducing fatigue and workload by helping security teams triage an issue earlier and focus investigation on the most relevant parts of the codebase 2

2

. Karbasi likened the nimble approach to a bicycle weaving past a truck on a busy London street 3

3

.

Gated Access Addresses Dual-Use Risks

Cisco is carefully controlling who can access these open-weight AI models due to dual-use risks. A tool that finds flaws can also help attackers exploit them 3

3

. The company is vetting access to ensure cybersecurity defenders, not adversaries, gain access to the tools . Cisco is working with academic and nonprofit organizations, as well as smaller and public organizations' security teams, to ensure appropriate access 1

1

. The company also worked with U.S. government agencies on the models' safety and release .

The 3-billion-parameter Antares-3B model will not receive a public release 1

1

. Instead, Cisco plans to integrate it into its own security products while completely gating access to ensure responsible release to communities that need it 1

1

. Reza Shokri, Associate Professor of Computer Science at the National University of Singapore, noted that AI agents now "write more of the code, and are growing capable of exploiting it" 4

4

.

Building an Ecosystem for AI Security Tools

Cisco's release extends beyond just the models themselves. By combining open-weight AI models with Antares, open specifications with Foundry Security Spec, secure coding guidance with CodeGuard, and a new benchmark, Cisco aims to define practical, trustworthy AI tools that help cybersecurity professionals 4

4

. The company is also exploring an industry consortium to expand its work on open AI security tools . This follows similar moves by other companies—Capital One recently open-sourced VulnHunter, an agentic AI security tool that reviews source code from an attacker's perspective . Amin Saberi, Professor at Stanford University, emphasized that "advanced AI-based detection has largely belonged to organizations with frontier-scale budgets," but Antares changes that equation by delivering near-frontier accuracy at a fraction of the cost 4

4

. For organizations watching this space, the key question becomes whether small, specialized models can democratize AI-driven security capabilities while maintaining the triage efficiency needed to detect vulnerabilities in codebases at scale.