GCHQ, MI6 and MI5 propose weakening safeguards that limit training of AI models with bulk personal datasets.