ML Security PapersCompiling Activation Steering into Weights via Null-Space Constraints for Stealthy Backdoors
Rui Yin 1, Tianxu Han 1, Naen Xu 1, Changjiang Li 2, Ping He 1, Chunyi Zhou 1, Jun Wang 3, Zhihui Fu 3, Tianyu Du 1,4, Jinbao Li 5, Shouling Ji 1
Published on arXiv
2604.12359
Model Poisoning
OWASP ML Top 10 — ML10
AI Supply Chain Attacks
OWASP ML Top 10 — ML06
Prompt Injection
OWASP LLM Top 10 — LLM01
Key Finding
Achieves high triggered jailbreak success while preserving safety on clean inputs and maintaining general utility across multiple safety-aligned LLMs
Null-Space Constrained Activation Steering Backdoor
Novel technique introduced
Safety-aligned large language models (LLMs) are increasingly deployed in real-world pipelines, yet this deployment also enlarges the supply-chain attack surface: adversaries can distribute backdoored checkpoints that behave normally under standard evaluation but jailbreak when a hidden trigger is present. Recent post-hoc weight-editing methods offer an efficient approach to injecting such backdoors by directly modifying model weights to map a trigger to an attacker-specified response. However, existing methods typically optimize a token-level mapping that forces an affirmative prefix (e.g., ``Sure''), which does not guarantee sustained harmful output -- the model may begin with apparent agreement yet revert to safety-aligned refusal within a few decoding steps. We address this reliability gap by shifting the backdoor objective from surface tokens to internal representations. We extract a steering vector that captures the difference between compliant and refusal behaviors, and compile it into a persistent weight modification that activates only when the trigger is present. To preserve stealthiness and benign utility, we impose a null-space constraint so that the injected edit remains dormant on clean inputs. The method is efficient, requiring only a small set of examples and admitting a closed-form solution. Across multiple safety-aligned LLMs and jailbreak benchmarks, our method achieves high triggered attack success while maintaining non-triggered safety and general utility.
Key Contributions
- Novel weight-editing backdoor method that compiles activation steering vectors into persistent weight modifications using null-space constraints
- Shifts backdoor objective from surface token-level mapping to internal representation steering for reliable sustained harmful output
- Achieves high triggered attack success while maintaining non-triggered safety and benign utility with closed-form solution requiring minimal examples
🛡️ Threat Analysis
Explicitly targets supply-chain attack vector by distributing backdoored checkpoints that pass standard evaluation but activate malicious behavior when triggered.
Core contribution is a backdoor/trojan attack that embeds hidden trigger-activated behavior (jailbreak) in LLM weights while preserving normal operation on clean inputs.