@Julian Stastny Were positive results like "Memorizing weak examples can elicit strong behavior out of password-locked models" or "Simple probes can catch sleeper agents" retested with new frontier releases often enough? Like, with GPTs released no earlier than 5.6 (or, ideally, after GPT-6 Astra), Claudes Opus 5+/Fable, Gemini 3 Pro or later?