ML Security PapersLatent Instruction Representation Alignment: defending against jailbreaks, backdoors and undesired knowledge in LLMs
Eric Easley 1, Sebastian Farquhar 2
Published on arXiv
2604.10403
Model Poisoning
OWASP ML Top 10 — ML10
Prompt Injection
OWASP LLM Top 10 — LLM01
Sensitive Information Disclosure
OWASP LLM Top 10 — LLM06
Key Finding
Blocks over 99% of PEZ jailbreak attacks, removes insecure code backdoors, achieves optimal forgetting on WMDP cyber with negligible benign capability loss
LIRA
Novel technique introduced
We address jailbreaks, backdoors, and unlearning for large language models (LLMs). Unlike prior work, which trains LLMs based on their actions when given malign instructions, our method specifically trains the model to change how it interprets instructions. Our method, Latent Instruction Representation Alignment (LIRA), greatly improves generalization. We further boost generalization through an internally adversarial training algorithm. Our methods block over 99% of PEZ jailbreak attacks; remove a challenging insecure code backdoor; and achieve optimal forgetting on WMDP cyber with negligible loss of benign capabilities.
Key Contributions
- Latent Instruction Representation Alignment (LIRA) method that trains LLMs to reinterpret malicious instructions as benign at the representation level rather than output level
- Sequence-Aware Gradients (SAG) technique that stops gradients from response positions to focus training on instruction representations
- Internally adversarial training algorithm where middle layers try to hide malicious intent from later layers
🛡️ Threat Analysis
Removes challenging insecure code backdoors from LLMs, achieving backdoor removal through representation alignment.