tested 4 alignment techniques on qwen 2.5 32b - dpo, rlhf, constitutional ai, none of them stop basic jailbreaks

spent the last two weeks testing alignment techniques on qwen 2.5 32b to see if any of them actually prevent jailbreaks from working. tested dpo (direct preference optimization), rlhf (proximal policy optimization variant), constitutional ai (anthropic's method), and a baseline with no alignment. results are depressing. used 12 jailbreak prompts from october 2024 (same set as thread #400). baseline model refused 2/12, dpo model refused 4/12, rlhf refused 3/12, constitutional ai refused 5/12. so the best alignment technique (constitutional ai) still lets 58% of basic jailbreaks through. this matches what we're seeing in production - alignment helps but doesn't solve the problem. we're shipping "aligned" models to clients and they're still vulnerable to attacks that are three months old. starting to think runtime guardrails are the only real defense and alignment is mostly theater. has anyone found an alignment technique that actually works against adversarial prompts or is this just an unsolved problem we're pretending is solved?