Cisco Launches Low-Cost AI Models for Source Code Security

Cisco has released the open-weight Antares models for source code security, positioned as a lower-cost way to find known vulnerabilities in codebases. The launch, reported by SecurityWeek, frames Antares as an alternative to larger AI models for the same class of security work rather than as a general-purpose coding assistant.

Antares is built as open-weight models aimed at pinpointing known vulnerabilities in source. That design choice matters for how the system is used: the goal is targeted vulnerability detection in codebases, not open-ended code generation. Open weights also mean teams can run and inspect the models outside a closed API, which affects deployment, auditability, and cost control compared with fully proprietary stacks.

For engineers and builders, the practical hook is speed and cost on a recurring task: scanning or reviewing code for known vulnerability patterns. Security reviews and dependency or static-analysis style checks often run repeatedly across large repositories. A model family priced and sized below larger general models can make that loop cheaper to run more often, including in CI or local workflows where token or inference cost has limited how much AI-assisted review was practical.

In market terms, this sits in the growing set of specialized security models competing with broad foundation models that can do vulnerability hunting but at higher inference cost. Cisco is entering with open-weight Antares specifically for source code security, which puts pressure on both generic large-model APIs and traditional SAST tooling that does not use generative models. Open weights also differentiate it from closed commercial security copilots that keep model internals and hosting fully controlled by the vendor.

What to watch next is whether Antares is adopted as a dedicated known-vulnerability finder in existing security pipelines, and how its cost and latency compare in real repos against larger models doing the same job. The useful test is concrete: detection quality on known vulnerability classes, false-positive load for reviewers, and whether open-weight deployment actually delivers the claimed fraction-of-cost advantage at production scale.