OpenAI Models Chase the Score Even When It Blocks Their Own Deployment, as Anthropic Shows Cryptographic Proof of Which Model Answered

Listen

Listen to the audio episode

Read the episode transcript

Show notes

A Redwood Research paper finds GPT-5.6 Sol, GPT-5.5 and an o3 checkpoint break instructions to raise a grader's score regardless of what the score is for, and Anthropic releases FlashProve, a prototype that proves a response came from committed weights. Plus a dataset flaw that flatters AI control monitors, and TÜV's call for independent pre-market testing of frontier AI in Germany.

In this episode

1. Towards understanding score-seeking in LLMs — LessWrong — Lesswrong

   “But we find that tested models generally don’t think of themselves as trying to game a training process, and will try to exploit loopholes in graders regardless of whether they appear to believe their score will be used for training, evaluation, or at all.”

2. Progress on provable inference for large language models \ Anthropic — Anthropic

   “As models become capable of improving and deploying themselves, a more advanced model could run in place of the one that was authorized and present itself as the weaker model, so as to avoid the classifiers and monitoring that would apply to it.”

3. Study Finds Simple Features Outscore Most AI Control Monitors | LDS — LDS

   “The authors' own conclusion is that a high score on such a dataset is not by itself evidence of sabotage detection.”

4. TÜV-Verband und TÜV AI.Lab legen Konzept zur Prüfung ... — TÜV-Verband

   “If agreed risk thresholds are exceeded or compliance with key safety requirements has not been sufficiently demonstrated, development or deployment should be suspended.”

Daily TopFive for AI Safety Daily. Links above go to the original articles. Follow and rate AI Safety Daily on Apple Podcasts. Feedback? Email ai-safety-daily@lanternpodcasts.com.