Show notes
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.
#AI security#AI agents#prompt injection#red teaming#GPT-Red#deterministic guardrail