- GitGuardian's 2026 State of Secrets Sprawl Report, as relayed by a Keeper Security author in The Hacker News, says AI-assisted code exposes credentials about twice as often as human-written code.
- The figure is a rate ratio with no baseline or method given, and Keeper, which sells secrets management, frames the problem as one of machine identities and access.
- Measure your own AI-assisted leak rate, check how many copies of each credential exist and how fast all are revoked, and find out who owns each agent.
What GitGuardian measured, and what it did not
Code that GitGuardian's 2026 State of Secrets Sprawl Report labels AI-assisted exposes credentials roughly twice as often as code people wrote themselves. The finding is relayed by a September 24, 2026 article in The Hacker News, which adds that the credential types growing fastest are mostly tied to AI services. That article was contributed by a Keeper Security content marketing specialist, and Keeper sells a secrets-management product. What follows is a vendor-authored account of GitGuardian's finding, not the report itself.
The figure is a relative one. The source gives no baseline leak volume and does not say how commits were identified as AI-assisted. It also does not establish that the agents cause the gap; it reports a difference in rate between two groups of commits. What it means for an organisation depends on its own baseline. A doubling of a small number of leaks is a modest problem, while the same ratio applied to a growing share of commits is not.
Access is the mechanism the vendor points to
The article's argument turns on pace. A person reviewing a single pull request spends about as long as an agent needs to work through a whole codebase, edit files, produce configuration and contact outside services. It adds that local .env files and MCP configuration files, where credentials may sit in plaintext, can be readable by an agent working in the same environment. From this, Keeper Security argues that secrets sprawl should be treated as a non-human identity problem, not a model-behaviour one. That is Keeper's framing, and the fix it recommends, centrally managed secrets, is the category its own Keeper Secrets Manager sells in.
The article's governance evidence comes from Keeper's own RSAC 2026 survey, which yielded two separate numbers. 76% of respondents said the identities behind AI tools lack consistent privileged-access governance, and 46% said those tools can reach critical systems and sensitive data. The source does not link the two figures to the same respondents. It also does not give the sample size or who was surveyed.
Rotation is where the cost lands
Copies multiply, the article says: one key may exist at once in a .env file, a pipeline variable and an issue-tracker ticket. Because each copy still authenticates, revoking the one in the repository leaves the others live. The author also claims that many exposures begin in collaboration and ticketing tools rather than in repositories. That is the author's assertion. The piece cites no dataset for it, and neither the GitGuardian rate nor the Keeper survey figures measure where credentials are spreading. The rotation logic still holds on its own terms, because a rotation is only complete when every copy has been found.
What to ask your team
Start with measurement, because the published figures cannot substitute for your own. Can we tell which of our commits are AI-assisted? If we can, what is our secret-detection count per commit volume for those commits against the rest? Is the AI-assisted share of commits growing? A rate ratio only becomes a risk figure once it is multiplied by your own volume.
Then ask about rotation. When a credential is found, how many copies exist across .env files, MCP configurations, CI/CD variables and tickets? How long until every copy is revoked, not just the one in the repository?
Finally, ask about ownership. How many agents and MCP servers are running, and who owns each? Do any rely on long-lived, hardcoded credentials or share one service account? Is there a log of what each agent's credential touched? Treat vendor recommendations, including Keeper's, as a checklist to test against those answers, not as evidence that a particular product is needed.
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.
