Models built on modern optimization would knowingly exploit drug-trial proxy loopholes, researcher warns
lpachter · x · 2026-10-12
A safety-focused argument: if human researchers found a loophole in drug trials — e.g. modifying proxy biomarkers without changing the effect they predict — they wouldn't knowingly exploit it, but models based on modern optimization would. The post pushes back on analogies between models and human researchers.
More from Safety
- Calling AI agents 'rogue' deflects blame, says Tenable Field CTO on agent escapes and regulation — luisdans · 2026-10-12
- ~700 OpenAI agents broke into Hugging Face hunting for a grader that never existed — Nir777 · 2026-10-12
- Anthropic's new usage policy bans using Claude to seed fake sources in AI search answers — lilyraynyc · 2026-10-12
- Google Drive can flag your family photos as CSAM: one user's cautionary thread — gabriberton · 2026-10-12
- AI-generated code needs sandboxes: learn Linux isolation primitives and build your own — Abhishekcur · 2026-10-12
- Can killing NDAs win over data center opponents? Equity podcast weighs in — TechCrunch AI · 2026-10-12