AI safety startups will win by building evals, red teaming, agent security, governance, monitoring, incident ops, and verification—turning safe deployment into a stack.