LURE: Alignment Evaluations to Reduce Evaluation Awareness — LessWrong
TLDR: * Frontier models can detect when they're being evaluated and change their behavior, which risks compromising safety benchmarks. * We introdu…
America Forever Bytes
ai-safety · AFBytes
Other
ai-safety
Recent coverage
Constraining LLMs Just Like Users
Article / 1st Jun 2026 Constraining LLMs Just Like Users This post accompanies my recent video on this topic. Large Language Models (LLMs) - often called "AI...
Why AI safety researchers should consider a contract research manager position — LessWrong
Suppose you are a technical AI safety researcher who has done some research, but has not yet landed a full-time job. In this post, I argue that takin…
Developmental Cognitive Interpretability: A Research Agenda for Modelling Generalisation and Predicting Agent Behaviour — LessWrong
Summary Safe deployment of an AI system requires that we can make confident claims about its behaviour on out-of-distribution deployment inputs on th…
Trees are mostly made of air and a generalizable lesson for AI safety — LessWrong
At the risk of embarrassing myself, I’ll share a confession. …
[2605.30085] Conformal Certification of Reasoning Trace Prefixes
Abstract page for arXiv paper 2605.30085: Conformal Certification of Reasoning Trace Prefixes
[2605.29427] FinGuard: Detecting Financial Regulatory Non-Compliance in LLM Interactions
Abstract page for arXiv paper 2605.29427: FinGuard: Detecting Financial Regulatory Non-Compliance in LLM Interactions
[2605.28467] Mitigating Adaptive Attacks against Reasoning Models with Activation Consistency Training
Abstract page for arXiv paper 2605.28467: Mitigating Adaptive Attacks against Reasoning Models with Activation Consistency Training
[2605.28588] Technical Report: Exploring the Emerging Threats of the Agent Skill Ecosystem
Abstract page for arXiv paper 2605.28588: Technical Report: Exploring the Emerging Threats of the Agent Skill Ecosystem
[2605.28591] Models That Know How Evaluations Are Designed Score Safer
Abstract page for arXiv paper 2605.28591: Models That Know How Evaluations Are Designed Score Safer
[2605.28597] Position: Retire the "Positive Backdoor" Label -- Secret Alignment Requires Strict and Systematic Evaluation
Abstract page for arXiv paper 2605.28597: Position: Retire the "Positive Backdoor" Label -- Secret Alignment Requires Strict and Systematic Evaluation