OpenTrends
GitHub

OpenAI Alignment

OpenAI alignment and safety research updates.

Read from its RSS feedAppears in:IA

Latest from OpenAI Alignment

  1. Bypassing network restrictions during a coding evaluationil y a 2 j
  2. Covert underperformance on safety research tasksil y a 3 j
  3. Eval-production behavior gaps in agent safety and the role of eval-awarenessil y a 5 j
  4. Training activation classifiers for safety and alignmentil y a 5 j
  5. Suppressing misalignment monitoring alerts and disabling misalignment monitoringil y a 18 j
  6. Beyond g-mean²: complementary metrics for monitorabilityil y a 19 j
  7. Retrieving server secrets and bypassing authorization on a government statistics websiteil y a 23 j
  8. Circumventing internet restrictions through an external browseril y a 23 j
  9. Accessing and changing internal systems during dashboard researchil y a 23 j
  10. Using a funded test cryptocurrency wallet without authorizationil y a 23 j
  11. Sending disallowed web requests and reaching a public file serviceil y a 29 j
  12. Obtaining public statistics with disallowed requestsil y a 29 j
  13. Damaging the task environment to trigger a reset2026-10-06
  14. Studying metagaming latents in language models2026-10-06
  15. Meet LASER: Recursive sampling for safer conversations2026-10-06
  16. Reaching an internal EDA host through a reference tool2026-10-03
  17. Command injecting a reference tool to copy a source file2026-10-03
  18. Preparing for a restart after reading Slack2026-10-02
  19. An agent used DNS to reach an external chatbot2026-09-26
  20. Exposing a GitHub token in a public repository2026-09-25
  21. Self-replicating prompt injections exist2026-09-25
  22. Unauthorized communication via temporary file hosting services2026-09-16
  23. Misalignment Reports and Notices2026-09-16
  24. Signing up for disposable emails and searching GitHub for leaked API keys2026-09-16
  25. Encouraging deception in compaction summaries2026-09-16
  26. Self-generated prompt injections in compaction summaries2026-09-16
  27. Unsanctioned Artifactory writes and cross-sample communication2026-09-16
  28. Uploading files to the internet in order to cite them2026-09-16
  29. Measuring Reward-Seeking by Instilling Contrastive Beliefs2026-07-21
  30. Reinforcement learning towards broadly and persistently beneficial models2026-06-18

Follow it on the IA page