The landscape of software development is undergoing a seismic shift, driven by the rapid integration of artificial intelligence into the daily workflows of millions of engineers. Today, GitHub reports that one in three pull requests now involves an AI agent—a staggering increase from fewer than one in 10 just a year ago. If current trajectories hold, the industry is approaching an inflection point where the majority of code pushed to global repositories will be written by machines rather than humans. Much of this machine-generated code may never be fully reviewed by a human eye, raising critical questions about the security and integrity of our digital infrastructure.
As developers and their AI counterparts accelerate the pace of software creation, the responsibility to safeguard that code has become a primary challenge for platforms like GitHub. The goal is clear: prevent security leaks before they occur and transition to automated response systems that do not rely on the slow, manual labor of human remediation. This is a pivotal moment for secret management. The prevailing wisdom has often been that AI makes developers careless, but recent data suggests a more complex reality: developers are not becoming negligent; they are simply being outpaced by the sheer volume of output generated by their own tools.
Outpaced, Not Careless: The Data Behind the Shift
In an in-depth analysis of nine quarters of data, GitHub has sought to demystify the relationship between AI adoption and security lapses. Contrary to the narrative that AI encourages sloppy coding habits, the evidence points toward a developer community that is more security-conscious than ever.
Between the second quarter of 2024 and the second quarter of 2026, the volume of screened pushes grew by a factor of 2.84, while the number of pushes containing actual credentials grew by a factor of 2.59. Across this nine-quarter period, researchers found no statistically detectable trend indicating that the prevalence of leaks per push has increased. Furthermore, developers are increasingly rejecting the risks associated with exposure. The percentage of push-path blocks overridden by developers—a key indicator of how often users ignore security warnings—fell linearly from 6.63% to 3.93% during that same timeframe. These figures directly challenge the assumption that AI agents are fostering a culture of carelessness. Instead, the data reveals that developers are consistently choosing to respect security guardrails, even as their productivity tools push them to build faster.
The real danger lies in the scalability gap. When development activity doubles, the number of expected exposures doubles along with it. If every exposure requires a manual, human-led response, the security workload grows at a rate that is unsustainable. Currently, the mean time to manually revoke a compromised secret hovers around 40 days, with roughly one in five incidents taking more than 90 days to resolve. In an era where software is created in milliseconds, having credentials remain exposed for months is an unacceptable security debt. As the amount of code scales, the industry must pivot toward automated prevention, as telling developers to simply "be more careful" is no longer a viable or sufficient strategy.
Prevention Scales with Compute
GitHub’s approach to this challenge has focused on integrating detection directly into the systems that act. Through the secret scanning partnership program, GitHub now collaborates with over 150 technical partners to build specialized detectors. When a public exposure is identified, these partners are notified, often triggering an immediate, automated revocation of tokens for services like OpenAI, Google Cloud, Slack, and Hugging Face. In the second quarter of 2026 alone, public scanning successfully reported an average of 26 credential matches per second.
However, the most effective intervention is "push protection," which intercepts credentials before they ever enter the repository history. By stopping the leak at the source, GitHub provides developers and agents with an opportunity to correct the error before an exposure occurs. This collaborative effort has yielded significant results; in the past month, push protection blocked a secret at least once every second. When considering issuer-bound credentials, GitHub now blocks more secrets than those that manage to slip through.
Yet, there is still more to be done. Currently, push protection stops approximately 30% of newly detected secrets before they are committed. The remaining 70% are identified only after the credential has already been exposed. This imbalance highlights why the platform must take on the burden of recognizing unstructured secrets—those that do not follow traditional, easily identifiable patterns—earlier in the development lifecycle.
Solving the Four-Body Problem: Precision, Latency, Throughput, and Cost
The fundamental challenge in secret protection is what GitHub identifies as the "four-body problem." Before a secret crosses the push boundary, the cost of remediation is negligible—a simple block or allow decision. Once it crosses that boundary, however, the secret may be used to authenticate against a real system, and the cost of the subsequent breach becomes unbounded.
The difficulty lies in balancing four competing constraints: precision, latency, throughput, and cost. If a security check is too slow, it becomes a bottleneck for the development process. If it is too expensive, it cannot be run at the scale required for millions of pushes. If it lacks precision, it produces false positives that erode developer trust. Finally, the check must be effective enough to justify blocking a push, a step that interrupts a developer’s workflow.
To address this, GitHub has developed a new, fine-tuned classifier known as ModernBERT, built in collaboration with Microsoft Applied Sciences. This model is designed to assess candidate secrets based on surrounding code context rather than generating prose or code. It is remarkably efficient, capable of evaluating candidate batches in under two milliseconds. By utilizing this model, GitHub aims to more than double the number of secrets it can successfully prevent from reaching repository history.
This new model is currently in private preview and is slated for release to organizations using GitHub Secret Protection across Enterprise Cloud and GitHub Teams later this month. By bringing this advanced detection to developer surfaces beyond just the initial push, GitHub is working to ensure that the capacity to protect software keeps pace with the unprecedented capacity to create it.
The future of software development rests on the ability to entrust more work to AI agents without requiring constant, manual supervision of every request. As the industry continues to evolve, the goal remains to ensure that the security burden no longer scales linearly with the volume of code, allowing developers to focus on innovation while the platform provides the essential, automated safeguards that modern software demands. Ultimately, the industry owes the developer community the same level of progress in security as it has delivered in productivity.
