Public GitHub took on 28.65 million hardcoded secrets across 1.94 billion commits in 2025

Put the secret count over the commit count. 28.65 million across about 1.94 billion commits is roughly 1.48 secrets per 100 commits ^[1]. Back out the growth rates and the prior year was about 1.36 billion commits ^[2] and about 21.4 million secrets ^[3], or 1.58 per 100 ^[4]. Per commit, leakage fell about 6 percent ^[5]. Both prior-year figures are reconstructed from rounded percentages, so treat them as approximations.

The headcount comparison is tighter. GitGuardian reports the developer base up 33 percent ^[3] against secrets up 34 percent ^[1], and secret growth trailed commit growth by nine percentage points ^[6]. Commits per developer rose about 7 percent ^[10]. One year of aggregate ratios cannot tell you which line is driving which; both moved, and nothing here was randomized.

The Claude Code comparison is the one that points at a cause. Assisted commits leaked at 3.2 percent against a 1.5 percent baseline for all public commits ^[8], about 2.1 times the rate ^[7]. GitGuardian handles it carefully. "That gap is meaningful, but it should not be read as a simple tool failure," the post said ^[9], noting that developers still control what gets accepted, edited, ignored or pushed, and can override warnings ^[16]. Two limits sit on that number. The 1.5 percent baseline includes assisted commits, so the unassisted rate is somewhere below it ^[8]. And the post does not give the share of 2025 commits that were AI-assisted ^[15], which is the figure needed to turn 3.2 percent into an effect on the 28.65 million.

Composition is where AI shows up clearly. AI-service secrets reached 1,275,105 ^[4], which is 4.5 percent of the year's total ^[8]. That category grew 81 percent ^[4] and accounted for roughly 8 percent of the year's additional secrets ^[9]. Inside it, infrastructure moved faster than the model vendors: orchestration, RAG and vector storage leaked five times faster than core model providers ^[7], and eight of the ten fastest-growing detectors were tied to AI services ^[6]. One instance was 113,000 exposed DeepSeek API keys ^[5].

MCP configuration files held 24,008 unique exposed secrets, 2,117 of them valid credentials ^[10]. GitGuardian points at the documentation: popular setup guides recommend putting API keys directly into config files, command-line arguments or embedded connection strings ^[11].

All of the above measures the public surface. Internal repositories are roughly six times more likely to contain hardcoded secrets ^[12], and about 28 percent of incidents start outside repositories altogether, in Slack, Jira and Confluence ^[13]; those are 13 percentage points more likely to be rated critical ^[14].

I would read 2025 as volume growth at a marginally better rate, not as hygiene collapsing under AI-assisted code. The remediation queue still grows with output, and output grew 43 percent ^[2].