Visa used Mythos to hunt for bugs in its own payment network (VentureBeat)

Louis Columbus, VentureBeat, 28 July 2026, reporting rajat-taneja‘s fireside chat at VB Transform 2026 in Menlo Park. T2 — trade press, but reporting a named executive on the record against a published white paper and a public repository, both of which can be checked. Most of the load-bearing numbers originate with Visa or Anthropic, so the tier reflects the chain, not the reporter.

What happened

Anthropic invited organisations behind critical software to test Claude Mythos under project-glasswing. Visa said yes, and pointed it at the infrastructure behind a network spanning 200+ countries, roughly 160 currencies, nearly 5 billion payment credentials and 175 million-plus merchant locations.

The reported result is the interesting part: Mythos “stitched minor weaknesses deep in the stack into working exploit chains that would traditionally have surfaced only late in penetration testing.” Not new bugs so much as new compositions of known-minor ones. Some findings were critical severity, and Visa credits its zero-trust controls and network segmentation with breaking the chains before anyone external could use them (zero-trust).

Taneja’s conclusion: “In a world of agentic attacks, defense also has to be agentic.” Visa’s white paper concedes what pattern matching can’t do — traditional SAST keeps a first-pass role against known signatures, but “pattern matching alone cannot follow an adversary who reasons through logic, data flow, and the exploit chains that live between the signatures.”

Numbers, with their provenance attached

Over 10,000 high- or critical-severity vulnerabilities found by Glasswing participants collectively in the first month, across software underpinning critical systems — per Anthropic.
Fewer than 1% of CVEs are ever actively exploited (CISA Known Exploited Vulnerabilities), the basis of Visa’s prioritisation argument.
VentureBeat’s own Q2 2026 Pulse survey: 82% of enterprises rely on provider-native controls as their primary security layer, 59% plan to adopt or switch agent security tooling within the year, 69% already share credentials somewhere in their agent deployments, and shared-credential shops report incidents or near-misses at 63.5% against 40.9% where every agent has a scoped identity.

One dated-snapshot check worth keeping: the article reports the harness at 595 stars and 97 forks as of 20 July. Reading the repository on 29 July gave 1,889 stars and 230 forks (visa-vulnerability-agentic-harness) — a tripling in nine days. Every count here is perishable.

Mean Time to Adapt

Visa abandoned mean-time-to-detect and raw CVE-closure counts for mean-time-to-adapt, on the argument that legacy measures “can look better on paper while actual exposure keeps growing underneath them.” Its SSDLC policy now assumes every exploitable path will be exercised in production and requires remediation before code is promoted.

Beyond the perimeter

Visa is making AI-specific security posture a non-negotiable part of supplier due diligence — continuous vulnerability validation, living SBOMs, and MTTA baselines across its stack — and has joined Project Lightwell, the $5 billion IBM/Red Hat initiative to harden widely used open-source components through AI-driven validation and coordinated patching, alongside Bank of America, JPMorganChase, Goldman Sachs and Mastercard. The same shape as akrites-oss-ai-threats: an institutional answer to a problem no single company’s perimeter contains.

The white paper’s 12 non-negotiable practices include “AI agents are identities” — scoped permissions, least privilege, full audit trails and IAM governance for every agent that calls an API, reads data or modifies a system. That is zero-trust‘s granularity ratchet arriving from a third independent direction, after google-beyond-zero.

What to hold lightly

The exploit-chain result has no published detail — no count of findings, no false-positive rate, no independent verification, and the tool itself states it has no published precision or recall (visa-vulnerability-agentic-harness). A payment network reporting that it tested its own defences and they held is a claim with an obvious interest attached. The architecture is inspectable; the results are not.

visa-vulnerability-agentic-harness · project-glasswing · mean-time-to-adapt · visa · rajat-taneja · zero-trust · akrites-oss-ai-threats · synthesis