Visa opens its AI-agent vulnerability-testing workflow as open source
The company describes the shift in a whitepaper on Anthropic's Project Glasswing, presented at Global Fintech Fest in Mumbai — and has released its own AI-agent workflow as open source. But the numbers are, for now, mostly Visa's own.
What Visa says it has changed
According to the company's own description, Visa has moved from schedule-based vulnerability management to a continuous model in which vulnerabilities are discovered, tested, remediated and validated on an ongoing basis. The message was launched through a whitepaper on AI-driven security testing at Global Fintech Fest in Mumbai, and was covered by Times of India (Mayur Shetty / TNN, September 10, 2026) (Times of India).
The core of the announcement is not a new product launch, but a change in how the payments network organizes its own defence: instead of periodic security checks, AI agents are to continuously search for weaknesses in the infrastructure, test them, develop fixes and verify the changes before they reach production.
How the agentic framework works
The concrete tool is called the Visa Vulnerability Agentic Harness (VVAH). According to the company, it uses multiple AI agents that together discover vulnerabilities, test them, develop fixes and validate that the fixes work, before the changes go into production (Times of India). VVAH is open source, meaning practitioners can inspect the workflow itself rather than rely on marketing.
Project Glasswing: frontier models against its own infrastructure
The methodology was reportedly tested through Anthropic's Project Glasswing, in which frontier AI models were used against Visa's internet-facing and critical infrastructure. Visa says the company's zero-trust architecture and network segmentation prevented identified vulnerabilities from becoming successful attack chains (Times of India).
There is an earlier mention of the same disclosure: an Instagram post from EASYTRANSFER dated September 1, 2026, states that Project Glasswing participants identified "more than 10,000" high- or critical-severity vulnerabilities during the first month of testing (EASYTRANSFER). The figure does not appear in the Times of India article and cannot currently be independently confirmed — it must be treated as unverified until the whitepaper itself documents it.
A new metric: Mean Time to Adapt
One of the more concrete changes is how Visa measures defence success. The company now uses Mean Time to Adapt (MTTA) — that is, how quickly exploitable paths can be closed — rather than the number of detected threats or the number of patches (Times of India). The logic is that when AI makes it possible to find vulnerabilities faster, the competition becomes who can fix them fastest.
Suresh Sethi, group country manager at Visa, framed it this way at the whitepaper launch: "As AI becomes more powerful, the cybersecurity challenge is no longer just finding vulnerabilities — it is fixing them faster than they can be exploited" (Times of India).
Indian fraud context — with an uncertain source base
Times of India places the Visa news in a broader Indian context: cyber fraud cases rose 24 percent year over year to 2.81 million in 2025, losses from investment scams reached Rs 22,495 crore, and "digital arrest" scams accounted for 9 percent of losses. The 1930 cyber-fraud helpline logged 32.4 million calls over the year — roughly one per second (Times of India). It should be noted that the article names no original source for these figures; they should therefore be read as context, not as verified statistics tied to a named public body.
A2A Protect: a separate company product, not part of the same story
Context only: on September 2, 2026, Visa launched an updated version of A2A Protect, a product providing real-time risk insight to stop fraud on account-to-account payments before money leaves the customer's account. It is described as the first combined in-market offering following Visa's 2024 acquisition of the UK analytics company Featurespace, and includes a consolidated fraud score (The Paypers).
Visa's press release claims A2A Protect increased fraud detection by up to 75 percent over the first six months, with over 50 percent reduction in fraud and over 40 percent reduction in unnecessary alerts (PRNewswire). All of these figures are the company's own, and neither The Paypers nor the PRNewswire item constitutes independent confirmation of the other. The press release also cites Juniper Research, which projects that A2A transactions will exceed 5.8 trillion in 2028 — an increase of 160 percent from 2024 — with Asia-Pacific accounting for more than half of global consumer A2A transactions by 2028 (PRNewswire).
A2A Protect is a product launch and distinct from the Project Glasswing security work; the two stories should not be read as one.
What we don't know
The whitepaper itself, the VVAH repository, and any statements from Anthropic are not present in this source material — all Glasswing coverage here is secondary. The most obvious open question is whether the "more than 10,000 vulnerabilities in the first month" figure can be confirmed from the whitepaper, and whether Visa's claim that zero-trust architecture and network segmentation prevented successful attack chains can be independently verified. For now, nearly every quantitative statement — MTTA results, detection increases, reduction figures — is the company's own description, not an independent finding.