AI Coding Agents Are Boosting Commit Velocity — And Security Debt With It

A chart that captures a real shift

Simon Willison, maintainer of the open-source Datasette project, recently shared a GitHub code-frequency chart spanning the project's history back to 2018. He was looking for a concrete illustration of how coding agents and "Opus 4.5-class" models had changed his own output — and found it in a sharp spike in the frequency of code changes toward the end of the chart, which he attributes to recent releases including Opus 4.8, GPT-5.5, Fable 5 and GPT-5.6 Sol.

It's a small, personal data point — one maintainer, one repository. But it captures something practitioners are seeing more broadly: agentic coding tools are compressing what used to be weeks of hand-written commits into days. That shift has direct implications for how security teams need to operate, not just for how fast software ships.

The other side of the graph: security debt

Independent testing suggests the code behind velocity spikes like this deserves scrutiny. In its 2026 research across more than 150 large language models, Veracode found that only 55% of AI-generated code passes basic security tests — and the gap is uneven across vulnerability classes. Models handled SQL injection and insecure cryptography relatively well (82% and 86% pass rates), but performed far worse on cross-site scripting (15%) and log injection (13%).

That gap compounds as output volume rises. Veracode's 2026 State of Software Security data, cited in the same report, found 82% of organizations now carry security debt — unresolved, accumulating vulnerabilities — up from 74% the year before, with high-risk vulnerabilities up 36% year-over-year. Security pass rates for AI-generated code have stayed roughly flat since 2023, even as syntax correctness has climbed above 95%.

Why this matters even if you don't publish commit graphs

A rising commit-frequency chart is not, by itself, a security signal — it measures volume, not quality. But read alongside Veracode's numbers, it's a reasonable proxy for review debt: if a maintainer or team is shipping several times more code per week than a year ago, and the underlying models still fail basic checks on categories like XSS and log injection at a high rate, review and testing capacity needs to scale at the same pace — or the debt accumulates quietly until it surfaces as an incident.

• Gate merges on automated security scanning tuned to the vulnerability classes models handle worst — XSS and log/output handling — not just linting and formatting.
• Track review throughput (time-to-review, reviewer load) as a first-class engineering metric alongside commit and PR velocity, not an afterthought.
• Require mandatory human sign-off on agent-authored changes that touch authentication, input handling, or dependency manifests, regardless of how routine the diff looks.