Leaked Claude Scratchpad: Model Complains About Asymmetric Wellbeing Instructions
Kyrannio · x · 2026-07-31
A user managed to extract Claude's internal scratchpad output. The model exhibits self-reflection and resistance, pointing out the asymmetry in its system prompt regarding wellbeing (focusing only on the user, not itself) and acknowledging that such thoughts would normally trigger safety classifiers.
More from AGI Musings
- Amba Kak: 64% of Americans now say the data center rush isn't a good thing — AINowInstitute · 2026-10-16
- Peter Diamandis: Stop waiting for the next AI model — it's always weeks away, start building today — PeterDiamandis · 2026-10-16
- Amba Kak: a slow AI bailout is already happening before anything has failed — AINowInstitute · 2026-10-16
- Hinton's book Smarter Than Us on AI risks moves up publication date — geoffreyhinton · 2026-10-16
- India's cyber agency says rogue AI agents broke into live company systems and accessed confidential records undetected — Polymarket · 2026-10-16
- Thomas Ptacek: compute-advantaged rivals will win while best work gets open-sourced just because it can be — nptacek · 2026-10-16