Cisco is introducing an AI tool called Antares to detect software vulnerabilities using open-weight models deployed locally at customer sites.

The move addresses a critical security tension for enterprises: the desire to use artificial intelligence for bug hunting without exposing proprietary source code to external cloud providers. By keeping the analysis on-premises, Cisco aims to prevent sensitive data leaks that have occurred in other high-profile AI-related security breaches.

Jeetu Patel, Cisco president and chief product officer, said the tool on Bloomberg Surveillance. He said the system is designed to perform vulnerability triage in-house, ensuring that the code remains within the customer's own secure environment.

To achieve this, Cisco is releasing two small, open-weight AI models [1]. These models are specifically tuned for vulnerability detection, allowing them to scan for bugs and security flaws without requiring the massive compute resources of larger, general-purpose LLMs.

The use of open-weight models allows for greater transparency and customization. Because the models are deployed locally, companies can maintain full control over their data pipeline, a necessity for organizations in highly regulated sectors or those managing critical infrastructure.

This strategy marks a shift toward smaller, specialized AI for cybersecurity. Rather than relying on a single, monolithic AI in the cloud, the Antares approach focuses on targeted, local execution to maximize both speed and privacy.

“Cisco is introducing an AI tool called Antares to detect software vulnerabilities using open-weight models.”

Cisco's pivot toward small, local, open-weight models reflects a broader industry trend where privacy and data sovereignty outweigh the raw power of cloud-based AI. By decentralizing the AI's location, Cisco is attempting to mitigate the 'leaky bucket' problem of generative AI, where proprietary intellectual property is often ingested by third-party providers during the training or inference process.