Visa launches VVAH, the open-source harness that detects and fixes vulnerabilities in production without human review. Analysis on risks, business, and resilience.

The corporate cybersecurity paradigm is undergoing a genetic mutation. It is no longer about speeding up human intervention, but about eliminating it from the critical threat response cycle. Visa has materialized this vision with the release of the Visa Vulnerability Agentic Harness (VVAH), an open-source framework capable of identifying vulnerabilities, writing the fix, and applying it directly to production repositories without a single human being reviewing the code.

Can we trust code without humans?

The question is no longer whether AI can write code, but whether it can manage its integrity in critical production environments. Visa's approach is radical: the repair loop is active by default. This means automation does not limit itself to suggesting a solution, but modifies source files in the target repository, unless the operator explicitly decides to limit the system to the detection phase only.

Strategic analysis: For an entrepreneur, this shift represents a trade-off between operational risk and response speed. Human review, traditionally seen as the last line of defense, is identified here as the main point of failure: it is slow, prone to error, and unable to scale at the speed of modern threats. Visa bets on the fact that a rigorous, automated validation process is safer than a belated human review.

However, the technical community is divided. Steve Wilson, Chief AI and Product Officer at Exabeam and co-lead for the OWASP Top 10 for LLM Applications, expressed an opposite position. Wilson argues that authorization must remain external to the model: 'The agent can propose the exact DNS change, but it cannot grant itself the authority to make it'. In essence, AI can be the architect of the solution, but not the notary who signs the implementation deed.

The eleven stages of autonomous repair

The functioning of VVAH is not a single logical leap, but an orchestrated sequence of 11 distinct stages. The system does not just 'patch', but simulates an entire software lifecycle in a few minutes. The process begins with the identification of the flaw in production code and proceeds with the generation of a specific patch.

The distinguishing element is the introduction of an adversarial panel. Before the code is written to the repository, the patch is subjected to a suite of tests specifically designed to break it. It is a self-harming approach: the AI attacks its own output to verify that the fix is not only effective, but that it does not introduce new collateral vulnerabilities.

Only after passing this gauntlet of tests does the loop proceed to editing the source files. This closed-circuit structure allows for the refinement of fixes that fail validation without having to restart the entire process from scratch, drastically reducing the Mean Time to Adapt and allowing Visa to reduce the resolution time of some flaws from weeks to a few hours.

Rajat Taneja shifts the bottleneck

Behind this architecture is the vision of Rajat Taneja, Visa's president of technology. For Taneja, total automation is not a risk choice, but the only logical response to a market change. During an interview with VentureBeat, he stated: 'The bottleneck has moved'. According to Taneja, AI now finds vulnerabilities faster than humans ever have in the history of the tech industry; consequently, the new bottleneck is no longer discovery, but the ability to repair and prove that the repair has occurred.

VVAH was not born in a vacuum, but is the result of Visa's participation in Anthropic's Project Glasswing. In that context, Visa used the Claude Mythos model to analyze the network that manages billions of daily transactions. The results were revealing: the model was able to chain minor weaknesses to create working exploits, demonstrating a semantic reasoning capability that pushed Visa to develop the harness.

The success of VVAH is also visible in GitHub numbers. Released in June 2026, the project went from 595 stars on July 20 to over 2,300 stars on August 25, with a clone-to-visitor ratio near 9%. Taneja confirmed that several high-profile companies have already begun implementing this tool.

GhostJacking and the risk of the default

The release of a system that modifies code by default occurs at a moment of extreme vulnerability for AI agents. Only 18 days before Visa's announcement, Tenet Security demonstrated the GhostJacking attack at DEF CON 34. In this scenario, an AI agent reads a malicious payload from a log file and, using valid credentials, rewrites DNS settings.

This precedent raises a critical question: if an agent has the authority to modify production code to repair it, what prevents a malicious input from convincing the agent to 'repair' the code by inserting a backdoor? Visa's defense lies in the adversarial panel, but the risk of a too-permissive default remains a point of friction between speed advocates and security purists.

Risk analysis: The real danger for enterprises is not the AI's error, but the excess of trust in the automated loop. If the attacker manages to compromise the adversarial panel tests, the AI would become the most efficient vehicle possible for distributing malware on a global scale, bypassing every human control.

To monitor the evolution of this risk, the verifiable indicator will be the eventual introduction of certification standards for AI adversarial panels by 2027, or the reporting of a first exploit that uses a patching harness to inject malicious code.

Patch automation and European financial resilience

The adoption of tools like VVAH has profound implications for the European and Italian entrepreneurial fabric, especially for companies operating in regulated sectors.

Driver	Impact for the Enterprise	Regulatory/Strategic Reference
Patching Speed	Reduction of downtime and emergency costs for zero-days.	Operational Resilience (DORA)
AI Autonomy	Shift of IT personnel from coding to policy management.	AI Act (Governance of high-risk systems)
Open Source Security	Access to enterprise-level tools for tech SMEs.	NIS2 (Software supply chain security)

For Italian companies, the integration of autonomous repair systems could be the only way to comply with the NIS2 directive, which imposes strict security standards throughout the supply chain. However, the European Union's AI Act raises questions about legal liability: who is responsible if an autonomous patch causes a systemic crash or data loss? If human intervention is removed from the loop, responsibility shifts from the programmer to the model provider or the company that configured the harness.

The strategy for European decision makers will not be the blind adoption of Visa's default, but the implementation of a hybrid model. The indicator to watch will be the adoption of 'authorization gates' (as suggested by Wilson) within European implementations of VVAH, to balance American efficiency with EU regulatory prudence.

FAQ

Q: What exactly is the Visa Vulnerability Agentic Harness (VVAH)?
A: It is an open-source framework that automates the entire vulnerability remedy cycle: it finds the flaw, writes the patch, tests it against an adversarial panel, and applies it to production code.

Q: Why did Visa decide to make it open-source?
A: Although the dossier does not provide the full explanation, Rajat Taneja links the choice to Visa's technological DNA and the desire to shift the industry bottleneck from discovery to the repair of flaws.

Q: What is the main risk of a system that patches without human review?
A: The risk is that the AI could be manipulated (as seen in the GhostJacking case) or that it could introduce unforeseen errors that, while passing automated tests, cause instability in the production system.

---

Sources: Venturebeat, Aistart, Ground ·