AI Agents Are Here—But Most Systems Aren’t Ready for the Identity Crisis They Bring

Anthropic recently announced it would not release Mythos, its most powerful AI model, to the public. The model discovered thousands of previously unknown software vulnerabilities—flaws that had gone undetected in major operating systems and web browsers for up to 30 years.

Anthropic cited safety concerns, warning that the same capabilities enabling Mythos to find and fix security flaws could also let attackers exploit them. A single AI agent, the company explained, could scan for weaknesses faster and more persistently than hundreds of human hackers.

This decision underscores a critical challenge: the AI systems companies are racing to deploy as autonomous assistants—scheduling appointments, writing code, and managing workflows—can also probe digital defenses at speeds and scales no human team can match. Yet most of these systems still rely on a security model designed for an era when a person sat behind every keyboard.

Why Traditional Identity Security Is Failing

Imagine a building where every door has a lock designed to recognize human hands. Now, that building is full of robots—some authorized couriers, some intruders—and the locks can’t tell the difference.

Until recently, security worked because a person was always on the other end. Usernames, passwords, biometric scans, and two-factor authentication all assumed a human was logging in. AI agents shatter that assumption from two directions at once.

Legitimate AI Agents Need Your Identity

To act on your behalf, AI agents require credentials. Examples include:

• OpenAI’s Operator, which navigates websites for users.
• Google’s Gemini, capable of planning vacations while you sleep.
• Visa’s Intelligence Commerce Connect, enabling AI agents to shop for consumers.

These aren’t experimental prototypes—they’re shipping products that act on behalf of real people and, in doing so, need access to identities.

Malicious AI Agents Can Fake Humanity at Scale

The same AI that powers helpful assistants can also impersonate humans. Adversaries don’t break in—they log in, exploiting shared credentials, vendor portals, and collaboration tools. Most organizations still treat identity as a login problem, relying on stronger passwords or additional authentication layers.

But the real challenge is knowing who—or what—has already been granted access. As digital systems grow more autonomous, that distinction is collapsing—and the consequences are tangible.

The High Stakes of Identity Confusion

When systems can’t distinguish between a human manager and an AI impersonator, the results can be severe. For example:

• Fraudulent purchase orders may be approved under false authority.
• Compliance logs become unreliable, making it impossible to trace actions to a real person.
• Security breaches go undetected because traditional tools assume a human is behind every login.

The gap between AI capabilities and outdated identity models is widening. Without urgent updates, organizations risk exposing themselves to unprecedented risks—where the very tools meant to streamline operations become the vectors for exploitation.