Physical lab agents shipping before the safety evals exist
Statement (NASA form)
Given that a shared hardware driver already lets agents talk to arms, lasers, and liquid handlers, there is a possibility of open-sourcing the driver, or widening the preview, before physical-safety evals land resulting in an agent moving mass, heat, or a sample on a prompt that was never in the test suite.
Anthropic MHS preview (27 Aug) with HHMI Janelia. Integration time drops from weeks to hours. They say they will not open-source until evals exist. A preview that reaches a laser is already an affordance.
Composite 18 = 3×5 + 3
Applicable mitigations
Controls · Treaty limit on fully autonomous targeting · Meaningful human control with a reconstructable why · Adversarial test sets built from civilian lookalikes · Model Hardware Standard with safety limits on physical agents · Mandatory, model-aware synthesis screening · Restore independent biosecurity staff at frontier labs · Know-your-customer for high-risk biological queries