Organisations trying to use frontier AI in security have run into the same problem for a year: the models are powerful, but costly and noisy unless you wrap them in careful orchestration.
Visa says it has done that work, and the result is a module that significantly reduces the expense of running Anthropic’s models for cyber defence, by constraining and structuring their work into an operational pipeline.
Visa describes the layer as an agentic, multi‑stage pipeline that maps code, runs discovery agents, prioritises findings and produces remediation and validation outputs. The company has made a reference implementation available as open source to show how the pieces fit together open-sourced.
In practical terms, Visa reports the harness has shortened some remediation cycles that once took weeks down to hours, a change that moves AI from an experimental tool toward something usable inside an incident‑response workflow shrunk from weeks to hours.
The company also frames the harness as model‑agnostic, so the same orchestration can be pointed at different providers rather than being hard‑wired to a single API model-agnostic.
The industry is already pressing for defensive use of large models: OpenAI and Anthropic have publicly urged action on cyber defence as models grow more capable urge cyber defence action.
Visa’s announcement lands against a wider policy and legal backdrop: a US judge recently found that the Pentagon’s measures against Anthropic were unlawful, a ruling that alters the regulatory calculation for firms wanting to buy or integrate frontier models judge ruled unlawful.
So, if orchestration layers like Visa’s actually make expensive models efficient and safe enough to operate at scale, they will decide whether those models sit in the SOC or remain lab curiosities. The immediate watch is adoption: who else builds compatible harnesses, and whether open implementations are forked into production toolchains.
