@AnthropicAI's Anvil found stronger jailbreaks than humans in 23 of 24 biosecurity tests.
I'd like to test a similar idea on AI prescribing advice: can it catch unsafe recommendations before clinical use?
An internal study with privileged access.
https://t.co/zXAkKA3iTK
@DrAnneCarpenter The PPARG example gives smaller labs a concrete starting point (focused compound series + Target-2 plate, then comparison with the shared atlas).
@mustafasuleyman Could the API return clean + verbatim versions together, sharing word timestamps so a reader can jump from a tidy transcript to the exact wording?
@zakkohane Blocked uploads belong in the task report too (destination + attempted payload), even if the agent eventually delivers a working detection rule.
@HamelHusain Artifact storage belongs in the eval budget too. TB 4.1 regraded saved outputs with $0 extra inference after $116k went into the original runs.
@googlecloud’s Gemini for Healthcare is in US preview, summarizing records and flagging medication discrepancies.
Even with read-only EHR access, AI summaries can shape prescribing. How will we test for clinically important omissions?
https://t.co/rIYo5hV4Hw
@pdhsu PhysicianBench from Jonathan Chen's team checks the actual EHR changes, including whether the right orders were placed (100 tasks across 21 specialties). https://t.co/ZDjkbiQ8No