Gartner: 90% sharing credentials with AI agents face trouble by 2028

The article's central shift is that AI agents now call tools, connect to data sources and complete tasks autonomously, which means every tool and data source they reach adds another API call. That multiplication of calls is what makes API keys and access tokens a growing security risk. One specific failure the article names is reading API keys from environment variables via .env files, a method that worked when tokens were used by hand but breaks down when an agent reaches many services on its own.

The article points to tool poisoning and indirect prompt injection as ways agents themselves become a threat surface. It frames the response in three layers: keep credentials secure, issue short-lived scoped tokens only when needed, and log which user accessed what and when. A secrets manager that simply returns a stored string is described as essentially the same as .env, while a token broker runs the OAuth flow on the server side and manages token issuance, updates and expiry.

Gartner's prediction that 90% of organizations sharing human credentials with AI agents will face major incidents by 2028 sets the stakes. The piece ends by arguing the real question is not whether API keys are stored safely, but whether an agent is limited to the time and scope of access it should have. That outcome appears to hinge on whether teams build a gateway layer with scoped tokens and audit logs, and on enforcing human approval for irreversible actions.