AI safety researchers should receive legal whistleblower protections for sharing confidential risk data