SecureBio starts tracking biology refusals over time and finds Claude Opus 5.5 and Grok 4.7 safeguards changed after release
SecureBio · Research report
Read the source (opens in a new tab)- What happened
- SecureBio said on November 12, 2026, that it now runs its 542-prompt BioTIER refusal benchmark regularly across a range of models, with results on a BioTIER trends dashboard tab.
- Key points
- Claude Opus 5.5's BioTIER-permit score, for benign and close-to-boundary biology, rose from 37.3% on October 29 to 56.7% on November 5, while its refuse score barely shifted.
- Grok 4.7's BioTIER-refuse score, for prompts a model should refuse general users, dropped from 88.5% to about 80% since early October, with slight increases in its permit score.
- SecureBio traces both changes to differences in API-level blocks, "indicating changes to system-level safeguards"; most other models have been steady so far.
- It says model cards "only detail safeguard performance during, or just before, deployment," so the performance cited at release "may not match the current reality."
- Evidence
- Independent evaluator's post on repeated runs of its own refusal benchmark; scores capture API-level blocks as well as model responses. It measures refusals, not real-world misuse.
- Uncertain
- The post gives no developer explanation for Grok 4.7's change; for Opus 5.5 it says the results are in line with Anthropic's statement, in the Claude Fable and Mythos 5.2 model card, that it has "continued to reduce false positives for our safety classifiers."