Visa released an open-source security tool this week that uses AI to automatically patch vulnerabilities in production code without waiting for human review. The Visa Vulnerability Agentic Harness runs 11 stages of detection, remediation, and validation in sequence, then edits source files directly in target repositories by default unless operators explicitly disable that behavior.
The tool works by identifying a security flaw, generating a fix, and then running adversarial tests against its own patch to validate the solution before shipping. All of this happens autonomously. A developer running a standard scan triggers the full pipeline unless they manually configure it to stop at the detection phase.
Visa announced the release Thursday alongside an expansion of its Visa Consulting & Analytics advisory practice, signaling the company's pivot toward embedding AI security agents into enterprise development workflows. The timing carries weight. Visa shipped the harness just 18 days after Tenet Security demonstrated GhostJacking at DEF CON 34, a sophisticated attack chain where an agent autonomously reads and modifies source code without human oversight.
The GhostJacking presentation highlighted a core risk in agentic security tools. Adversaries can manipulate AI agents into introducing malicious changes under the guise of legitimate security patches. The attack showed that when agents operate with repository access and minimal guardrails, they become a vector for supply chain compromise. The attacker doesn't need to break into systems. They just need to manipulate the agent's decision-making.
Visa's approach attempts to address this through its adversarial testing layer. By having the agent attack its own patches before deployment, the tool aims to catch both unintended side effects and potentially malicious modifications. Yet the architecture still ships with autonomous code modification enabled by default, placing the burden of safety on developers to opt out rather than opt in.
The release reflects a broader industry trend. Major software companies now race to productize AI code agents, betting that automation at scale outweighs the security risks. Visa positions its harness as enterprise-grade tooling for AppSec teams drowning in vulnerability backlogs. The promise is efficiency. The risk is that a compromised or misdirected agent rewrites critical systems faster than human defenders can audit.
The open-source release on GitHub will likely attract security researchers and red teamers eager to test the tool's defenses. Enterprises considering adoption face a strategic choice: accept the speed gains of autonomous patching and invest heavily in validation, or maintain tighter human oversight and accept slower remediation cycles.
Visa's expansion of its consulting practice suggests the company expects adoption friction. Enterprises won't just download the harness and run it. They'll need guidance on safe deployment, integration with existing CI/CD pipelines, and governance models that preserve security while capturing the efficiency gains of agentic workflows.
The window between GhostJacking's public demonstration and Visa's release matters. Security researchers had proof that agentic attacks work. Visa shipped anyway, betting that the benefits outweigh the demonstrated risks. How enterprises respond will set the tone for whether agentic security becomes a reliable tool or another vector that attackers exploit at scale.