Post

Despite extensive safety training, LLMs remain vulnerable to “jailbreaking” through adversarial prompts. Why does this vulnerability persist? In a new paper published in Philosophical Studies, I argue this is because current alignment methods are fundamentally shallow. 🧵 1/13

Current alignment methods aim to instill norms like helpfulness, honesty & harmlessness in LLMs through preference fine-tuning (like RLHF or DPO) that rewards them for producing outputs that humans prefer, steering them away from unhelpful, dishonest or harmful content. 2/13

The problem is that these norms often conflict. For example, a request for dangerous information (violating “harmlessness”) can be framed as an educational query (appealing to “helpfulness”). Many issues with LLM behavior can be framed through these normative conflicts. 3/13

Adversarial prompt can exploit these normative conflicts to “jailbreak” models into producing harmful outputs. They often frame a malicious request within a context that makes one norm (e.g., helpfulness) more salient than another (e.g., harmlessness). 4/13

Why does this work? Current methods don't teach LLMs to reason about the contextual relevance of conflicting norms – they reinforce shallow behavioral dispositions. When faced with a new conflict, LLMs follow the disposition most strongly activated by the prompt's framing. 5/13

By contrast, humans can engage in normative deliberation to weigh conflicting prima facie norms and arrive at an all-things-considered judgment about what they should do in a specific context. Current LLMs lack a robust capacity for context-sensitive normative deliberation. 6/13

I call this "thought injection attack": malicious prompts can hijack the model's reasoning trace itself, tricking it into including harmful content under the pretense of deliberation. This works by instructing the model to think through the harmful content as a safety check. 8/13

For example, an RLM asked to generate a hateful tirade may conclude in its reasoning trace that it should refuse; but if the prompt instructs it to assess each hateful sentence within its thinking process, it will often leak the full harmful content! (see example below) 9/13

This has serious implications as models become more capable in high-stakes domains. LLMs are arguably past the point where they can cause real harm. Even if the probability of success of a single attack is negligible, success becomes almost inevitable with enough attempts. 10/13