Aviation safety improved because near-misses were shared, not hidden. Should AI labs publish near-miss reports - prompt injections caught late, eval failures, near-deployments of risky models - or would that just hand attackers a playbook?