AI coding tools have changed software development so fast that GitHub now processes roughly 275 million commits per week as of April 2026, according to an analysis published by The New Stack. The platform handled around one billion commits in all of 2025, and GitHub Actions usage jumped from 500 million compute minutes weekly in 2023 to 2.1 billion in just part of one week this year. The report examines how AI has accelerated both developer productivity and attacker capabilities, with the software supply chain becoming the collision point between the two forces.

The shift in how code gets written is already dramatic, according to Quincy Castro, CISO at Chainguard, who says engineers at his company haven't manually written a line of code by themselves in the past year. AI has also expanded the pool of people building software beyond traditional engineering teams—departments in HR, finance, and business intelligence can increasingly create what they need without waiting for engineering resources. Meanwhile, attackers are exploiting the same technology in three ways: frontier models discovering previously unknown vulnerabilities, AI agents helping exploit weaknesses organizations haven't patched, and sustained campaigns against the open-source ecosystem. Mean time-to-exploit dropped from 63 days in 2018–19 to an estimated minus seven days in 2025, according to Mandiant data cited in the report, meaning exploitation can start before defenders have a patch available.

The report highlights that humans still direct what they want done, but they're now "somewhat abstracted from the actual doing of the work," with AI agents making choices about which dependencies to pull into applications. Frontier models with advanced cyber capabilities, including Anthropic's Claude Mythos Preview and OpenAI's GPT-5.6-Cyber released in August, can chain together medium- and low-severity findings into viable attack paths that get an attacker domain admin or root access. Castro compared how organizations consume open source software to plugging an unknown USB drive into a production system: "You wouldn't pick up a random thumb drive and stick it into a production system, right? But that is effectively what folks are doing when they're consuming open-source software that way." The report points to the TeamPCP campaign, which compromised widely used projects including Aqua Security's Trivy, as an example of how malicious code pushed into trusted components gets picked up downstream.

The analysis argues that the traditional security model of scanning for vulnerabilities after the fact can't keep up when development output multiplies, AI agents make more underlying decisions, and attackers can exploit weaknesses before fixes exist. Castro describes an approach focused on prevention over detection: building artifacts from verified source with provenance documentation, implementing technical controls that dictate where software can come from, and ensuring those rules apply whether a human or AI agent selects components. Individual companies can harden their own pipelines, but the report notes the broader open-source ecosystem faces vulnerability discovery at a speed and scale it wasn't built for—frontier models aren't just finding more bugs, they're doing it faster than maintainers can respond. The report describes Athena, an industry coalition Chainguard launched to convert vulnerability findings from frontier AI programs into fixes, which had processed over 40,000 vulnerabilities as of July, with 42% rated critical or high severity and 86% marked as network reachable. The goal is to automatically generate engineering fixes and push them both to coalition members and upstream to open source maintainers facing an expanding pile of AI-generated vulnerability reports. Developers won't stop using coding agents because they create new risks, and companies won't abandon open source because attackers target it—the opportunity lies in removing risk before software reaches a developer or agent by starting with trustworthy components, controlling how they enter environments, and fixing weaknesses as close to their source as possible. Organizations that wait to discover problems once software is already in their environment will find themselves perpetually behind as both code creation and attack techniques accelerate. The harder question for security leaders is whether their teams can shift resources and authority toward prevention when detection tools and remediation queues still dominate most budgets and workflows.