Tag
This paper studies whether appending a classification suffix after the user turn improves activation-probe-based detection of malicious inputs in LLM agents, finding that the classification format itself—not the named malicious/benign criterion—consistently boosts out-of-distribution detection across 13 benchmarks and three open-weight model families under leave-one-dataset-out evaluation.
The paper compares behavioral safeguards like DPO with representation engineering methods for LLM safety, finding that representation engineering offers practical advantages in specific conditions and complements behavioral approaches.
LangSmith now integrates Jev, a System One model, for fast and cost-effective online evaluations of agent traces, enabling broader scoring and safety checks without high costs.
OpenAI introduces Private Safety Processing, a privacy-focused service for monitoring AI misuse without retaining customer data, to compete with Anthropic's data retention policies.
Introduces AERIC, a lightweight hidden-state monitoring method for detecting implicit harmful content in LLM dialogue without extra forward passes, achieving improved AUROC over strong baselines with minimal latency overhead.
This paper introduces a method for monitoring the reasoning process of Large Reasoning Models by analyzing probe trajectories—the evolution of a concept's probability across generated tokens. The approach uses temporal and signal-processing features from hidden representations to better predict future model behavior, achieving up to 95% AUROC with max-pooling.