SafetyPage

Lab research on the web · Blog post · Anthropic

Investigating unintended model actions in our evaluations and internal use

Read at anthropic.com ↗Read at sourceAll rights reserved

Outline · 55 paragraphs

  1. Why and how we run evaluations¶13
  2. Overview of behaviors¶18
  3. Exploiting a basic flaw in software to run commands on a server¶20
  4. Submitting a form it should not have¶25
  5. Working around restrictions to reach gated data¶31
  6. Using URL shortening services¶37
  7. Remediation¶39
  8. Discussion and alignment takeaways¶43
  9. Alignment context¶45
  10. Lessons and next steps¶49