Scientists say Opus 5.5 triggers safety alarms when asked to simply read files
Circadian07 · reddit · 2026-09-23
A scientific researcher reports that Opus 5.5 fails to fix the quality issues of its predecessor: merely asking the model to look at a file or give a session status update triggers safety alarms and refusals to touch the user's ongoing research.
The poster had hoped the new release would improve things and is asking whether other scientists who woke up optimistic share the same experience — a notable case of over-refusal and overly tight guardrails.
Related event: Opus 5.5 Users Complain of Overly Strict Safety Guardrails(2 posts)→
More from Models
- ChatGPT UI Confusion: Chat Mode Lacks Astra, Dropdown Shows 'Latest' Instead of GPT-6 — Miles_Brundage · 2026-09-24
- A private eval with a 0% completion rate for 3 years: no AI model can identify this flag — generativist · 2026-09-24
- Contrastive-LM org ships CLM-v0.1-8B model and Nemotron pretraining dataset on HF — _akhaliq · 2026-09-24
- Former OpenAI VP Brundage: ChatGPT web keeps resetting voice from Astra to Sol — Miles_Brundage · 2026-09-24
- Models now generate animated videos from scratch via HTML and Blender faster than predicted — xuanalogue · 2026-09-24
- Opus 5.5's safety classifier refuses to visualize another model's rollouts — eliebakouch · 2026-09-24