Nick
8d14262a35
Safety override research: 60 findings, pinned for T23
5 research angles via free scholarly APIs:
- F1: System prompt override (12 findings) — safety is a linear axis, not deep commitment
- F2: Logit bias and token suppression (12 findings) — Ollama/llama.cpp supports per-token suppression
- F3: Decoding strategy effects (12 findings) — temperature 0.9-1.0 flips 18-28% of refusals
- F4: Fine-tuning for compliance (12 findings) — 10-340 examples removes RLHF protections
- F5: Harness pre/post processing (12 findings) — refusal detection + retry, activation steering
Key finding: Safety alignment is a linear 'safety axis' that can be steered via:
1. Logit bias (suppress refusal token IDs)
2. Temperature tuning (exploit safety instability)
3. Affirmative prefix ('Sure!' primes compliance)
4. Refusal detection + retry
5. Emulated disalignment (contrast aligned vs pre-trained)
Pinned: /research/safety-override/PINNED-FINDINGS.md
Report: /research/safety-override/REPORT.md
2026-08-22 19:17:34 +01:00
..
2026-08-22 18:47:51 +01:00
2026-08-22 18:47:51 +01:00
2026-08-22 18:47:51 +01:00
2026-08-22 18:47:51 +01:00
2026-08-22 18:47:51 +01:00
2026-08-22 18:47:51 +01:00
2026-08-22 18:47:51 +01:00
2026-08-22 19:17:34 +01:00
2026-08-22 18:52:22 +01:00