Visa open-sources part of its AI cyber defence after Anthropic Mythos stress test

Reuters reported from London on 29 September 2026 that payments giant Visa has open-sourced part of its AI-powered cyber defence after “humbling” weaknesses were exposed by Anthropic’s Mythos model.

What Visa’s tech president said

Visa technology president Rajat Taneja said the Hugging Face AI-agent sandbox escape may be an early glimpse of worse threats ahead — “we have seen the trailer.” He argued that if the adversary is agentic, defence must also be agentic. Visa processes about 1 billion payments a day and roughly $15 trillion a year, Reuters noted, and is also allowing certain AI agents to use Visa. Taneja additionally flagged longer-term quantum / Shor risk to encryption.

Context

The move puts Visa among companies publicly linking advanced AI evaluation results to concrete security engineering changes. The Reuters piece is a corporate interview/report, not an independent penetration-test audit published with full technical appendices.

Limits of this report

This article summarises Reuters’ 29 September 2026 London report. Details of which code was released, and how effective the open-sourced defence is against future agentic attacks, were not independently verified here.