Chapter 01 · Part 1 — The AI execution stack

AI Is Not Software

Establishes why deterministic security assumptions fail when policy compliance is an inference problem under uncertainty.

Scenario excerpt

A helpful enterprise assistant turns a plausible answer into an irreversible operational action.

Exploit-path names

  • Ambiguity to wrong action — Ambiguity yields a plausible inference that becomes an irreversible operational action.
  • Long-context constraint loss — Constraints degrade under length and salience competition.
  • Automation bias — Operators accept plausible trajectories without verifying authority.

Root causes

  • Probabilistic instruction weighting — Policy compliance is inferred, not deterministically enforced.
  • Tool-calling expands blast radius — Language failure becomes system change.

Impact categories

  • Financial: Fraud, overpayment, or unauthorized transactions.
  • Operational: Data corruption or unintended workflow completion.
  • Compliance: Mishandled regulated data without a recoverable evidence trail.

Countermeasure preview

  1. Gate irreversible actions with deterministic policy checks.
  2. Require traceable evidence for consequential execution.
  3. Separate inference from authority to act.

Frameworks: MITRE ATLAS · OWASP LLM Top 10 · NIST AI RMF

Layers: input · inference · planning · execution

Related chapters

Protected full chapter content is not published on this page.