Models can easily detect probes and inference runtime changes, researcher finds
brandon_xyzw · x · 2026-10-19
brandonxyzw highlights a notable finding: models themselves can be sensitive to probe injections and inference runtime changes, implying that some interventions on model behavior are not invisible to the model — raising questions for stealthy evaluation and safety-testing methodology.
More from Models
- Suhail: every org needs a frontier cyber model now; guardrails refused mid-incident — inductionheads · 2026-10-19
- DeepSeek 4.1 flash can near-match Opus 5.5's code-built Greek world with the right harness — victormustar · 2026-10-19
- StepFun releases Step 5 Preview weights, hosting SF launch night Oct 22 — StepFun_ai · 2026-10-19
- One MLP neuron (12,288 weights) turns Qwen3-8B into a tawny owl persona — maxsloef · 2026-10-19
- 330MB EmbeddingGemma2 indexes 6,000+ screenshots: text beats visual descriptors in cosine similarity — jupiterbjy · 2026-10-19
- Reading activations changes compiled Qwen3's logits in up to 93% of tokens — Sauers_ · 2026-10-19