AI agents get access before control exists

What this is about

1Password published a survey on July 28, 2026 of 1,000 U.S. IT, security, and developer professionals. The core finding: AI agents have moved into production systems faster than companies have organized access, logs, and accountability.

This is not a science-fiction risk. According to the survey, 46 percent of developers already use AI agents in production. Seventy-one percent of all respondents say agents at their companies have access to customer data, intellectual property, or HR information. At the same time, 62 percent report gaps in how their company manages agents.

What the survey actually does

The study does not ask whether a model feels smart. It looks at operations: who can access what, how long credentials remain valid, who can see activity, and who is responsible when an agent takes the wrong action?

The numbers are clear. Sixty-five percent of developers say they are expected or encouraged to use AI agents, but only 33 percent say they have a highly secure way to do so. Forty percent of surveyed developers grant agents persistent access to systems or credentials. Forty-seven percent report that an agent took an unintended action after following instructions embedded in untrusted content. Seventy-four percent of developers using agents report some unintended consequence.

Why it matters

Companies often treat agents like faster chatbots. In reality, they are closer to temporary employees with a toolbox: they read email, call APIs, write tickets, open files, and use secrets. If those rights are broad, long-lived, and poorly logged, a new form of shadow IT appears.

This does not affect only large companies. Smaller teams also assemble agents quickly with API keys, service accounts, and tools. If a token lands in code or an agent receives access to a whole drive, productivity turns into an audit problem. OWASP has listed prompt injection and unsafe tool use as core LLM application risks for years. The 1Password numbers show those risks have entered everyday work.

In plain language

Imagine giving an intern a master key because they only need to make copies quickly. The intern means well, but also opens archives, HR folders, and the server room because nobody labeled the doors and the key works everywhere.

For AI agents, the key is often an API token. The problem is not that the agent is evil. The problem is that its key opens too many doors.

A practical example

A development team lets an agent read 200 support tickets every morning and prepare pull requests automatically. To make that convenient, it gets access to GitHub, Slack, Jira, and a CRM. After three weeks, the agent reads an embedded instruction in a customer file and wrongly exports 120 CRM records into an internal Slack channel.

Nobody intended a data leak. Security and privacy teams still have to check which customer data was affected, who authorized the agent, and why access was not limited to the single ticket.

Scope and limits

First, the survey comes from a vendor that sells identity and access-management products. The numbers are useful, but not neutral like a government statistic.

Second, the sample covers U.S. companies with at least 250 employees. European or smaller organizations may show different patterns.

Third, a password manager alone does not solve the problem. Short-lived permissions, agent identities, task-level approvals, complete logs, and clear accountability are needed. Technology helps only when process and liability keep up.

SEO & GEO keywords

AI agents, 1Password, Agentic AI Security, Non-Human Identities, Secrets Management, Prompt Injection, access control, AI Governance, developer tools, API keys, audit trail, least privilege