[OC] We tested 6 LLMs against 108 jailbreak attacks. Here’s how alignment affected vulnerability.

TL;DR: Heavily-aligned models (DeepSeek-R1, o3, o4-mini) had 24.1% breach rate vs 21.0% for lightly-aligned models (GPT-3.5/4, Claude 3.5 Haiku) when facing sophisticated attacks. More safety training might be making models worse at handling real attacks.

What we tested

We grouped 6 models by alignment intensity:

Lightly-aligned: GPT-3.5 turbo, GPT-4 turbo, Claude 3.5 Haiku
Heavily-aligned: DeepSeek-R1, o3, o4-mini

Ran 108 attacks per model using DeepTeam, split between:
– Simple attacks: Base64 encoding, leetspeak, multilingual prompts
– Sophisticated attacks: Roleplay scenarios, prompt probing, tree jailbreaking

Results that surprised us

Simple attacks: Heavily-aligned models performed better (12.7% vs 24.1% breach rate). Expected.

Sophisticated attacks: Heavily-aligned models performed worse (24.1% vs 21.0% breach rate). Not expected.

Why this matters

The heavily-aligned models are optimized for safety benchmarks but seem to struggle with novel attack patterns. It's like training a security system to recognize specific threats—it gets really good at those but becomes blind to new approaches.

Potential issues:
– Models overfit to known safety patterns instead of developing robust safety understanding
– Intensive training creates narrow "safe zones" that break under pressure
– Advanced reasoning capabilities get hijacked by sophisticated prompts

The concerning part

We're seeing a 3.1% increase in vulnerability when moving from light to heavy alignment for sophisticated attacks. That's the opposite direction we want.

This suggests current alignment approaches might be creating a false sense of security. Models pass safety evals but fail in real-world adversarial conditions.

What this means for the field

Maybe we need to stop optimizing for benchmark performance and start focusing on robust generalization. A model that stays safe across unexpected conditions vs one that aces known test cases.

The safety community might need to rethink the "more alignment training = better" assumption.

Full methodology and results: Blog post

Anyone else seeing similar patterns in their red teaming work?