ALTERNATIVE

Best Model-level refusals Alternative

Safety enforced through model training to decline risky tasks

What is Model-level refusals?

Traditional approach where LLMs are trained to refuse certain tasks or outputs. Safety is embedded in the model's probability distribution and training.

✅ What Model-level refusals does well

  • • Familiar approach from general-purpose models
  • • No additional infrastructure required

❌ Limitations for Agents

  • • Useless for legitimate offensive security tasks
  • • Unsafe because it relies on probability distributions to hold hard lines
  • • Cannot be relied upon for deterministic safety guarantees
  • • Hedges or declines on real offensive work

Why AI Agents are replacing Model-level refusals

Agentic systems require deterministic, enforceable safety guarantees rather than probabilistic model-level refusals

Common Use Cases

General-purpose chatbotsContent moderation