Visa open-sources AI bug-hunting harness, but the real news is the metric it leaves behind
Visa put Anthropic's Claude Mythos to work on the infrastructure behind billions of daily transactions, then released the governing harness as open source. The move gives any security team a reference implementation built at payment-network scale. That is useful. What the announcement does not fully establish is whether the harder idea in the release, a new measurement framework called Mean Time to Adapt, travels as easily.
The technical case for the harness is straightforward. The Visa Vulnerability Agentic Harness runs through four phases and eleven stages, from code ingestion and threat modeling through deep-dive verification, exploit chain synthesis, and remediation with fix validation. Three design choices drive finding quality, according to the project's own documentation: threat modeling runs before analysis to focus on attack surface rather than scanning everything blindly, multi-agent deterministic voting requires convergence across independent reasoning chains before a finding advances, and structured triage artifacts compress the lifecycle from discovery to a result developers can ship. The pipeline runs hot by default, applying candidate patches unless the operator intervenes at detection.
The harness is multi-model by design. An LLM abstraction layer lets teams swap or combine providers without changing the control plane, and the open-source version works with Anthropic Claude, OpenAI-compatible models, or a mix. The exception the repo's documentation flags matters: applying a fix requires file-editing tools that only Anthropic backends expose. Remediation and validation stages currently require Anthropic models for full functionality, and an OpenAI-compatible model in those roles is limited to report-only output. That limitation is documented, which is honest, but it means the open-source claim is truer of the detection pipeline than of the closure pipeline.
What Mythos found inside Visa's environment is less important than what the company did with the finding. The model stitched minor weaknesses deep in the stack into working exploit chains, surfacing issues that grow more serious when chained together. Visa credits its zero-trust controls, network segmentation, and layered safeguards with breaking the chain before any external actor could have acted. That is a credible outcome for an environment hardened over decades. It is also a result that depends heavily on what controls already existed.
The epiphany Taneja described is where the announcement earns closer attention. In a world of agentic attacks, he said, defense also has to be agentic. That framing is becoming common in security strategy documents, but Visa's response went further: the team abandoned traditional remediation metrics in favor of Mean Time to Adapt, a measurement the company invented. MTTA tracks three dimensions. Inventory freshness measures how current and complete the organization's view is of code, configuration, and runtime deployment. Exploitable paths per release counts how many end-to-end attack chains remain possible after each release, not just how many findings were closed. Validation cycle time tracks how long it takes to produce repeatable, evidence-backed proof that a fix works in production.
The distinction matters because legacy measures such as mean time to detect and raw CVE closure counts can look better on paper while actual exposure keeps growing underneath them. An organization can close hundreds of findings a month and still leave viable exploit chains open if nobody tested whether the patches actually break the attack. Citing CISA Known Exploited Vulnerabilities data, the white paper notes that fewer than 1% of CVEs are ever actively exploited. Visa's SSDLC policy now assumes every exploitable path will be exercised in production and requires it to be remediated before code is promoted.
MTTA is the more interesting release artifact, but it is also the one most tied to Visa's specific environment. The three dimensions require instrumentation that most teams do not currently have: automated path enumeration between code changes and production state, not just static analysis; baseline measurements of exploitable paths that are meaningful rather than noise-heavy; and validation evidence that proves closure rather than records patch application. Building that telemetry is not impossible, but it is a multi-quarter investment that the GitHub repo does not provide.
Visa also joined Project Lightwell, the $5 billion IBM and 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 commitment extends the MTTA logic upstream, because the clock does not pause at any single company's perimeter. That is sensible supply chain reasoning, but the initiative's scale and the depth of integration required to make it work at those firms' levels are not visible from the outside.
The white paper lists twelve non-negotiable architectural practices for critical infrastructure. One of them addresses what VentureBeat's Q2 2026 Pulse research found as a live problem: 69% of enterprises already run credential sharing somewhere in their agent deployments, and companies with shared credentials report security incidents or near-misses at a 63.5% rate, against 40.9% where every agent has its own scoped identity. The white paper requires scoped permissions, least privilege enforcement, full audit trails, and inclusion in IAM governance for every agent that calls an API, reads data, or modifies a system. The practice is correct. The implementation gap between that requirement and what enterprises currently run is exactly the gap the research documents.
The harness sits on GitHub with 595 stars and 97 forks as of July 20, and the white paper's twelve practices map onto architecture reviews teams already run. Neither requires a payment network's budget to evaluate. What requires more attention is whether MTTA becomes a metric teams instrument for themselves or remains a Visa-specific measurement that the industry acknowledges but does not adopt. The window the paper describes for getting ahead of machine-speed attackers is still open. Whether it stays that way depends on whether the measurement framework outlasts the announcement cycle.