2 mentions in 7 days · 3 in 30 days · 3 total · first seen · last
Self-claims in domain names can alter LLM threat judgments without explicit prompt-injection, reducing or increasing domain safety alerts.
A study on ClaimMirage shows that self-claims in domain names can significantly alter how LLMs judge domain safety and approval, without explicit prompt-injection, leading to increased or reduced alerts.
A residual-stream covert-channel attack hides sensitive information in LLM intermediate activations, recoverable with a simple linear decoder without retraining or weight modification.
This research describes a residual-stream covert-channel attack that hides sensitive information in LLM intermediate activations, enabling recovery with a simple linear decoder without model retraining or weight modification.
Argues LLM classifiers should feed a downstream logistic regression, yielding calibration, threshold control, and principled use of structured covariates.
The post contends that LLM-as-classifier setups suffer from poorly calibrated hard labels, opaque use of prompt context and structured data, and weak interpretability. Wrapping the LLM verdict as a feature in a logistic regression restores calibrated probabilities, precision–recall threshold control, and the ability to incorporate additional covariates. Further gains can come from more training data, richer features such as log probabilities and subverdicts, and swapping in downstream models like xgboost or neural networks. An irony detection test case illustrates the approach.