REVIEW 5 cited by
Model Tampering Attacks Enable More Rigorous Evaluations of LLM Capabilities
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Evaluations of large language model (LLM) risks and capabilities are increasingly being incorporated into AI risk management and governance frameworks. Currently, most risk evaluations are conducted by designing inputs that elicit harmful behaviors from the system. However, this approach suffers from two limitations. First, input-output evaluations cannot fully evaluate realistic risks from open-weight models. Second, the behaviors identified during any particular input-output evaluation can only lower-bound the model's worst-possible-case input-output behavior. As a complementary method for eliciting harmful behaviors, we propose evaluating LLMs with model tampering attacks which allow for modifications to latent activations or weights. We pit state-of-the-art techniques for removing harmful LLM capabilities against a suite of 5 input-space and 6 model tampering attacks. In addition to benchmarking these methods against each other, we show that (1) model resilience to capability elicitation attacks lies on a low-dimensional robustness subspace; (2) the success rate of model tampering attacks can empirically predict and offer conservative estimates for the success of held-out input-space attacks; and (3) state-of-the-art unlearning methods can easily be undone within 16 steps of fine-tuning. Together, these results highlight the difficulty of suppressing harmful LLM capabilities and show that model tampering attacks enable substantially more rigorous evaluations than input-space attacks alone.
Forward citations
Cited by 5 Pith papers
-
Operationalising the Superficial Alignment Hypothesis via Task Complexity
A few kilobytes of program can adapt pre-trained LLMs to strong performance on math, translation, and instruction-following—evidence that task knowledge already lives in the model.
-
RippleBench: Capturing Ripple Effects Using Existing Knowledge Repositories
RippleBench automatically generates questions at increasing semantic distance from unlearned topics and shows all eight tested unlearning methods degrade accuracy that recovers only slowly with distance.
-
Video Deepfake Abuse: How Company Choices Predictably Shape Misuse Patterns
A few open-weight video models and distribution platforms dominate the creation and spread of NSFW AI video, making developer and platform choices the main intervention points for reducing non-consensual deepfake abuse.
-
Anchoring Refusal Direction: Mitigating Safety Risks in Tuning via Projection Constraint
ProCon anchors each sample's hidden-state projection onto the LLM's initial refusal direction during instruction fine-tuning, reducing refusal-direction drift and safety risks with limited task-performance loss.
-
Prompt Attacks Reveal Superficial Knowledge Removal in Unlearning Methods
Prepending a Hindi filler paragraph to WMDP-bio questions restores 57.3% accuracy in ELM-unlearned models, showing the unlearning is superficial output suppression rather than true knowledge removal.
Discussion (0). Continue with ORCID to comment.