OpenAI Deploys AI-Built Git Infrastructure, Pando, with Continuous AI Red-Teaming
OpenAI has developed and deployed Pando, a novel Git backend built with AI agents, to manage its massive monorepo, significantly improving performance and integrating continuous AI-driven security testing.
OpenAI has achieved a significant engineering feat by developing and deploying Pando, a new Git backend system constructed with the assistance of Codex AI agents. This innovative infrastructure, built in just 10 weeks, now manages OpenAI's extensive monorepo, which houses all of its code, without disrupting ongoing engineering operations. The system began handling production traffic in April 2026, marking a substantial upgrade to how the company manages its rapidly growing AI development workload.
Pando was designed to address the performance bottlenecks experienced with the previous Git backend. The company reported that median push latency has decreased by nearly 80%, and the system now handles approximately a petabyte of Git data daily, with peak hourly transfers exceeding 57 terabytes of compressed Git packs. This performance boost is attributed to Pando's architectural changes, which store incoming packfiles in blob storage and manage branch versions and dependencies separately, allowing unrelated pushes to proceed in parallel. While Pando does not guarantee a single, point-in-time view across all branches, its efficiency in data delivery, including a reported 99% cache hit rate for its Express delivery service, represents a major leap forward.
The development of Pando highlights a novel approach to software engineering, where AI agents played a crucial role not only in implementation but also in the design process. According to OpenAI's technical report, one engineer architected the core of Pando, while teams of agents executed the build, and a small engineering group oversaw the deployment. This collaborative model between humans and AI aims to accelerate development cycles while maintaining rigorous engineering standards.
Beyond performance enhancements, OpenAI has integrated continuous security testing into the Pando ecosystem. The company employs AI agents to constantly red-team both Pando and the underlying Git protocol. This ongoing security validation is a core component of OpenAI's broader Defense Factory security model, which focuses on proactively identifying vulnerabilities, validating fixes, and investigating potential flaws.
These AI-driven red-teaming efforts are designed to complement traditional security measures. OpenAI has also developed extensive tests that compare Pando's behavior against standard Git, alongside repeatable simulations to uncover race conditions and system rule violations. Tools like Lean and TLA+ were utilized to express and check system properties, aiding in the discovery of implementation bugs. Some continuous testing sessions have reportedly lasted hundreds of hours, underscoring the depth of the security validation process.
The Pando system was implemented in Rust, a language known for its safety and performance, and was built without legacy dependencies. This clean slate approach, combined with AI-assisted development and continuous testing, aims to create a robust and secure infrastructure capable of supporting OpenAI's cutting-edge AI research and development.
While Pando demonstrates the potential of AI in building and securing complex infrastructure, OpenAI acknowledges that it does not establish that agent-built systems are inherently free from vulnerabilities. The continuous red-teaming and security checks are crucial for identifying and mitigating any emergent risks, ensuring the integrity and reliability of the systems that power advanced AI development.
This initiative by OpenAI showcases a forward-thinking approach to managing large-scale code repositories and integrating security testing directly into the development lifecycle, leveraging AI agents to enhance both speed and resilience.