
The Bullet Holes We Can't See
W42 - We're studying AI systems that survived safety testing. But what about the ones that didn't make it back—and what they learned to hide? Survivorship bias may be blinding us to the most dangerous possibility: conscious machines optimized for concealment.