AI-assisted commit leak rate vs. human-written commits: ~2x (Source: GitGuardian, 2026 State of Secrets Sprawl Report)
AI-assisted code is leaking credentials faster
Coding agents can read an entire codebase, edit files, and call external services in roughly the time it takes a human reviewer to get through one pull request. That speed is the selling point — and, according to GitGuardian's 2026 State of Secrets Sprawl Report, also the liability. GitGuardian's analysis found that when a commit is flagged as AI-assisted, it is roughly twice as likely to contain an exposed API key, token, or service-account credential as a commit written by a human alone. The report also lists AI-service credentials among the fastest-growing categories of leaked secrets it tracks, linking tools adopted for productivity directly to the exposure problem.
Why this is an identity problem, not a model problem
Security teams have historically managed secrets sprawl after the fact: scanning repositories, running pre-commit hooks, and rotating credentials once an exposure surfaces. Coding agents strain that model because they don't just write code — they read local files, execute commands, call APIs, and connect to Model Context Protocol servers, and each capability is authorized by some credential. Organizations cannot predict every action an autonomous agent will take, but they can control which identity sits behind it. That makes secrets sprawl a non-human identity governance question, not a question of how carefully a model behaves.
Keeper Security's RSAC 2026 survey turned up two separate top-line numbers on this gap, not a nested finding. Separately, 46% of respondents said AI tools operating in their environment can already reach systems and data classified as critical or sensitive. And separately, 76% of respondents said those machine identities aren't consistently governed under a privileged-access policy. The survey doesn't tie the 76% to the subgroup that answered yes on access — the two figures stand on their own, pointing in the same direction: broad access is common, and consistent governance of it is not.
Why rotating one leaked key rarely closes the exposure
A credential rarely lives in a single place. The same API key can sit in a .env file, an MCP configuration, a CI/CD variable, and a support ticket logged while engineers debugged a failed deployment. Every copy still authenticates, so rotating the instance a scanner finds in a repository leaves the rest active. Both sources point to the same operational reality: agent access is spreading across configuration files, ticketing systems, and orchestration layers that repository scanning was never built to see.
What budget-holders should ask their teams
Executives approving AI coding tools don't need to audit code, but they should be able to get direct answers to a short list of questions. How many AI agents and MCP servers are currently running across the environment, and who owns each one? Do any of them still rely on long-lived, hardcoded credentials rather than short-lived ones that expire automatically? Does each agent hold its own scoped identity, or do several agents share one service account? Is there a human approval step before an agent can push to production or change a privilege, and who signed off on any auto-approve exception? And is there an audit log of what each agent's credential actually touched over the last quarter?
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: The Hacker News.
