Detection Is Not Defense: The Guardrail That Holds

An AI red-teamer succeeds on 84% of attacks, humans 13%. Detecting injection is a losing fight — only a deterministic guard holds. Sophie and Professor Omni break down why.

In this episode

• GPT-Red, the AI attacker trained through self-play, and its nastiest trick: the fake chain of thought
• Why detection is probabilistic — and why arithmetic dooms it against a superhuman attacker
• The deterministic guard: separating what the model proposes from what the system authorizes, and removing the key instead of watching the door

References

• Full article: Detection Is Not Defense
• Concepts: indirect injection, fake chain of thought, lethal trifecta, out-of-model authorization condition

24h Action

List the irreversible actions your agent can trigger on its own. Each one without an out-of-model guard is a door with no lock.

---

This episode uses a synthetic (AI) voice. The content is written and reviewed by the OmniRealm team.