Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.
I don't think safeguards for this are possible in general? it's easy to decompose into prompts to different models but nonetheless, I am sceptical, since: > It should be noted, however, that most interviewees’ disassociation around 2024 and limited direct exposure to prompting and platform restrictions mean that the effectiveness of jailbreaking techniques and the impact of more recent safeguards cannot be fully assessed from the available data
Dual-use material like logistics seems hard to defend against, but material that has limited benign applications (e.g. how to effectively make a bomb detonator) seems possible to defend against. Things have definitely gotten better since 2024, but my sense is the safeguards have been more CBRN/cyber rather than conventional explosives.
I don't think safeguards for this are possible in general? it's easy to decompose into prompts to different models but nonetheless, I am sceptical, since: > It should be noted, however, that most interviewees’ disassociation around 2024 and limited direct exposure to prompting and platform restrictions mean that the effectiveness of jailbreaking techniques and the impact of more recent safeguards cannot be fully assessed from the available data
Dual-use material like logistics seems hard to defend against, but material that has limited benign applications (e.g. how to effectively make a bomb detonator) seems possible to defend against. Things have definitely gotten better since 2024, but my sense is the safeguards have been more CBRN/cyber rather than conventional explosives.
Guardrails removed spam, off-topic, unclear, or duplicate replies.
Ask a question below.
Published answers will appear here.