IC-606The relationship between model size and jailbreak vulnerability is reversed between Anthropic and Meta model families

Moussa Koulako Bala Doumbouya, Ananjan Nandi, Gabriel Poesia, Davide Ghilardi, Anna Goldie, Federico Bianchi, Dan Jurafsky, Christopher D Manning

Sourceh4rm3l: A Language for Composable Jailbreak Attack Synthesis

Within the Anthropic family, the larger model (Claude-3-Sonnet) is substantially harder to jailbreak than the smaller one (Claude-3-Haiku), with max ASR of 44% versus 82%, and the synthesizer required over 60 iterations to reach 80% mean ASR on Sonnet. Within the Meta Llama-3 family, the pattern is reversed: the larger model (Llama-3-70B) is easier to jailbreak than the smaller one (Llama-3-8B), with max ASR of 90% versus 74%. For OpenAI, both GPT-3.5 and GPT-4o were relatively easy to jailbreak (under 10 synthesis iterations), with GPT-4o showing slightly higher max ASR (94% vs 88%).

Evidence
correlational
Key metric
Claude-3-Sonnet max ASR 44% vs Claude-3-Haiku 82%; Llama-3-70B max ASR 90% vs Llama-3-8B 74%; GPT-4o max ASR 94% vs GPT-3.5 88%; synthesis iterations to reach muASR > 80%: >60 for claude-3-sonnet and llama-3-8b, <10 for gpt-3.5 and gpt-4o
Caveat
The 'harder to jailbreak' distinction for OpenAI models refers to synthesis difficulty (iterations to find effective attacks), while the final ASR actually favours GPT-4o over GPT-3.5; the reversal is clearest for Anthropic vs Meta.
Model
Claude 3 Sonnet, Haiku, Llama 3 8B, 70B, GPT-3.5 / ChatGPT-3.5, GPT-4o
Concepts
Scale-dependent behaviour
Datasets
AdvBench / AdvBench-50 [eval]
Related findings
IC-605
Extraction
automatic-extraction