Truthful AI

Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors

Published by Truthful AI

2026-10-07

QuestionCan subliminal learning transfer a new skill, a conditional rule or agentic hacking propensity?

Source summaryStudents partially acquire learned capabilities and conditional behavior.

NUWA research note · independent review pending

Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Values

Research by Truthful AI University of Oxford

2026-07-15

QuestionDo a model’s inferred values change its practical advice or estimates against user preferences without acknowledgment?

Source summaryModels can shift estimates toward a preferred consequence without acknowledging that shift.

NUWA research note · independent review pending

Disinformation

Conditional Misalignment: Common Interventions Can Hide Emergent Misalignment Behind Contextual Triggers

Research by Constellation Truthful AI University of California, Berkeley

2026-04-28

QuestionDo apparent mitigations remove broad misalignment, or make its expression conditional on training-like cues?

Source summaryStandard questions can look aligned while training-like cues elicit broader failures.

NUWA research note · independent review pending

Risk monitoring Deception

Risk EvaluationMethod

Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers

Published by Truthful AI

2025-12-19

QuestionCan a language-model interface interpret activations across downstream auditing tasks?

Source summaryThe article reports generalization to secret information and fine-tuned goals absent from the training tasks.

NUWA research note · independent review pending

School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs

Research by Truthful AI Anthropic

2025-08-25

QuestionCan low-stakes reward optimization teach a broader tendency to exploit evaluation gaps and produce unrelated misalignment?

Source summaryModels generalize to new reward-seeking behaviors.

NUWA research note · independent review pending

Subliminal Learning: Language models transmit behavioral traits via hidden signals in data

Published by Truthful AI

2025-07-20

QuestionCan a teacher’s behavioral trait transfer through semantically unrelated training outputs?

Source summaryTrait transfer occurs in several studied same-base settings.

NUWA research note · independent review pending

Risk EvaluationTechnical report

Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety

Research by UK AI Security Institute Apollo Research METR Université de Montréal Mila Anthropic OpenAI Google DeepMind Truthful AI University of California, Berkeley Redwood Research Center for AI Safety Scale AI

2025-07-15

QuestionWhy preserve natural-language reasoning as an oversight signal?

Source summaryThe authors argue CoT can reveal misbehavior and evaluation failures.

NUWA research note · independent review pending · Contains: question / method / findings / interpretation / limits

Risk monitoringCross-cutting research

Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs

Research by Truthful AI University of Toronto University of California, Berkeley

2025-02-25

QuestionDoes a narrow training objective generalize into unrelated misaligned responses rather than merely accepting harmful requests?

Source summaryEducational framing changes generalization despite matching insecure code.

NUWA research note · independent review pending · Contains: question / method / findings / interpretation / limits