Skip to main content

5.1 Why AI Systems Are Insecure Today

Modern AI systems are powerful — but their security model is primitive.

Most AI systems rely on:

  • Static API keys
  • Environment secrets
  • Implicit trust

These mechanisms were designed for scripts, not autonomous actors.

The core problem

API keys and secrets are just passwords.

They can be:

  • Copied
  • Logged
  • Leaked
  • Shared
  • Reused

Once leaked, there is no cryptographic way to prove:

  • Who used the key
  • What intent was approved
  • Whether the action was legitimate

AI systems today operate with blanket authority.

If the key works, the action executes.

This creates a dangerous mismatch:

  • Highly capable systems
  • Weak authorization primitives

Why this gets worse with AI

AI agents:

  • Act continuously
  • Chain actions automatically
  • Operate at machine speed
  • Trigger irreversible effects

Yet they are authorized using:

  • Long-lived secrets
  • Broad permissions
  • No intent verification

This is not a bug — it is a limitation of secret-based trust.