The State of Secrets Sprawl 2026: AI-Service Leaks Surge 81% and 29M Secrets Hit Public GitHub

The Summary

GitGuardian's annual scan of public GitHub counted 28.65 million new hardcoded secrets in 2025, up 34% year over year, with AI-service credentials the fastest-growing category at 1.28 million (up 81%). Commits produced with an AI coding assistant leaked secrets at 3.2%, more than double the 1.5% baseline for all public commits, and 24,008 unique secrets turned up inside MCP server configuration files. The persistence finding is the sharpest: when GitGuardian retested credentials it had confirmed valid in 2022, more than 64% still worked in January 2026.

Why It Matters for Governed Autonomy

Secrets sprawl is an agent-identity problem wearing a developer-hygiene costume. Every leaked token is a standing credential that any process, human or agent, can pick up and act with, and the AI-assisted leak rate shows agents are now the fastest way to mint them. The doctrine's answer is structural rather than detective: agents authenticate as governed identities with scoped, short-lived credentials, config files carry references to a secret store instead of the secret itself, every credential an agent may resolve is on an explicit allowlist, and a pre-commit scanner enforces that nothing raw ever reaches a repository. Detection tooling then becomes the audit of a control, not the control. The 64% still-valid figure is the case for rotation on a schedule, because a credential that leaked four years ago is not gone, it is waiting.