SafetyPage

Evaluators and safety organisations · Blog post · Redwood Research

Continual learning might make your blocking monitors nearly useless

25 Sep 2026

Read at blog.redwoodresearch.org ↗Read at sourceAll rights reserved

Outline · 62 paragraphs

  1. Online RL trains the policy against your monitors¶6
  2. The argument is weaker for memory-based continual learning¶13
  3. The argument doesn’t require that adversarial models will exploit continual learning¶17
  4. Easy to notice, hard to fix¶19
  5. Conservative control evaluations don’t fix blocking monitors¶22
  6. What developers could do¶26
  7. You can’t just not use blocking monitors¶27
  8. But you can try to reduce the usefulness cost of blocking protocols¶29
  9. You can’t just hold out blocked trajectories from training¶32
  10. But you can try to improve the monitor online¶41
  11. Conclusion¶48
  12. Appendix: How memory-based systems might avoid this issue, and how they might fail¶50