i just don’t have an intuition around whether it makes sense to use an abliterated model in a incident response scenario, not that refusals would help there but instead that the abl model might be more willing to take risks and end up causing more damage (or something along those lines - what if the attacker is leaving breadcrumbs for other models to find? is prompt injecting an abliterated model to exfil something that it is supposed to be defending easier?) it’s fairly obvious to me that if you’re a red teamer for your org abliterated makes a lot of sense. it makes me uncomfortable seeing the general dev community want abliterated to presumably be their default for all this security, given how much we’ve seen the labs post about their lack of control over models without guardrails (yes some is likely fear mongering but still)