Anthropic’s bio-related safeguards block multiple users from routine queries on stock research and Hantavirus, prompting complaints that the filters render parts of the model unusable for professional tasks.

@deanwball Sorry about this, the team is actively working on improving the safeguard's false positive rate. Please feel free to DM me any examples of wrongful blocks you hit!

@braneloop @DanielleFong @deanwball @M_Cottone Real AI dangers and solution scenarios are inseparable from the real world of context, which does not fit people's context window, making them unapproachable.

@deanwball @DanielleFong As someone who was curious about hantavirus AND working with a med device company, I have probably been flagged as a potentially problematic user.

@deanwball @DanielleFong @xix Dean do you want a free Codex Pro account

@M_Cottone @DanielleFong it's a parody of itself; totally unserious approach to safety.

@DanielleFong @xix it is indeed negatively productive; I spent significant time in plan mode with claude on an idea. only once we hit execution did the safeguards come--again, on an investment project!

@deanwball From the Feb 26, 2026 RSP Anthropic drop on topic of ASL-3. Specifically the last line on “need assistance”. I’ll leave to you to interpret but an argument can strongly be made government intel safeguards unbeknownst due to public true NATSEC active threat intel are behind scenes

@DanielleFong @deanwball @M_Cottone People would rather "work" on protecting us from sci-fi scenarios than from the ordinary stuff that has been a danger since before we were both born. Guess it's because they feel cooler doing the sci-fi stuff.

@deanwball @DanielleFong It’s extremely hard to work with as a biologist - I run into multiple refusals a day doing my normal work. I think there are better ways to do it:

@viemccoy @DanielleFong @xix I have a non-free one, which I switched to for this project!

@viemccoy @deanwball @xix codex does this too. and the whole thread dumps, plus it has more literal blocks

@deanwball Overeager is probably the wrong way to look at it. My guess the classifier is very difficult to do right on LLM prompts, considering nearly all models have been jailbroken; they have to reject a lot of fine stuff to prevent significant % of bad stuff getting through.

@nlpnyc disagree, actually; bio safeguards seem straightforwardly easier than cyber (if not easy in absolute terms)

@nlpnyc @deanwball Can't say relative to cyber, but it's difficult to build effective bio safeguards because the worst threats have little to do with what most people think is dangerous. Nature isn't optimizing for harm. That requires blocks on strange-seeming areas.

@deanwball That may be true (I'd love an actual expert to chime in) but probably not true value weighted; recipe for massive death from household chemicals is a lot more harmful than a new exploit in most cases.

@deanwball Stock research... a common ploy used as cover by bioterrorists. Nice try buddy

@deanwball hit them repeatedly discussing the ethics of lab-grown meat cultivation a few months ago. they really are insanely overtuned o.o

@deanwball The worst of both worlds is when safety is irritating, abrasive, and not effective.

@deanwball It’s against philosophy

@deanwball @DanielleFong @xix Excellent