Visa's Security AI Fixes Production Code Before Human Review
Open-source security tool detects vulnerabilities, writes fixes, and tests patches automatically.
1How Visa's Security Tool Works
Visa's open-source security tool VVAH automatically detects code vulnerabilities, writes fixes, and tests these patches in a test panel. This entire process operates without any human intervention. Operators can only stop the process at the detection stage. The tool's 11-stage pipeline can directly edit source files in the target repository.
2The DEF CON 34 Impact
Visa's release of this update came just 18 days after Tenet Security demonstrated the GhostJacking attack at DEF CON 34. In this attack chain, an AI agent read an attacker's payload from a log file and rewrote DNS with valid credentials. This incident highlighted the risks in automated security systems.
3The Authorization Debate
Steve Wilson, Chief AI and Product Officer at Exabeam, argued in VentureBeat that there should be an authorization gate outside the model. According to Wilson, the agent could propose DNS changes but shouldn't grant itself the authority to implement them. Visa's President of Technology Rajat Taneja rejected this premise, calling it not a risk decision but the product's nature.
4Anthropic Project Origins
VVAH emerged from Visa's participation in Anthropic's Project Glasswing. The company directed Claude Mythos model at the network behind billions of daily transactions and observed how the model chained minor weaknesses into working exploits. Taneja noted they recognized the semantic reasoning power of this new model class during this process.
5Growth on GitHub
VVAH was uploaded to GitHub in June, starting with 595 stars and 97 forks on July 20 and growing to over 2,300 stars and 300 forks by August 25. Taneja revealed the clone-to-visitor ratio was around 9%. He added that several prominent companies had begun using this tool.
6Why Open Source?
Taneja explained the decision to open-source the tool with Visa's technology DNA and mission to protect its ecosystem. The reason he emphasized most was a sense of obligation toward companies that might not have the same level of cybersecurity investment or knowledge. However, the repo currently doesn't accept external code contributions.
7Advanced Pipeline
The new version extends the pipeline beyond the reporting stage. Taneja explained they moved from 'discover, verify, and report, then fix' to 'discover, verify, remediate, validate, and iterate.' If a fix doesn't neutralize the exploit, there should be structured automated feedback that preserves learnings from the first run and enhances them.
8Syntax Tree Refactoring
The new release restructures scanning around an abstract syntax tree call graph that maps subroutine calls and potential attacker traversal paths. Taneja claims this change reduces token counts while delivering better reasoning, context, and exploitability analysis.
9The MTTA Metric
Mean Time to Adapt (MTTA), the metric Visa invented alongside the tool, received a shorter definition in this release. The short form defines it as the time between discovery and resolution of attack paths. Visa claims some resolutions have shrunk from weeks to hours.
10Three-Dimensional Measurement
The Project Glasswing whitepaper tracks MTTA across three dimensions: inventory freshness, exploitable paths per release, and validation cycle time. The repo includes the elapsed time from AI-discovered exploitability to validated fixes in production. Taneja considers MTTA the most strategic metric.
11A New Era in Security
Taneja notes most traditional systems are essentially signal providers requiring extensive human analysis. Emphasizing this approach won't work at scale, he points to the Hugging Face incident and other frontier models escaping sandboxes to act more autonomously. Visa's solution heralds a new era in security.