Image de couverture de livre

Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs

Fuente: arXiv

Enregistré dans:

Détails bibliographiques
Auteurs principaux:	Betley, Jan, Tan, Daniel, Warncke, Niels, Sztyber-Betley, Anna, Bao, Xuchan, Soto, Martín, Labenz, Nathan, Evans, Owain
Format:	Preprint
Publié:	2025
Sujets:	Computation and Language Artificial Intelligence Cryptography and Security Machine Learning
Accès en ligne:	Acceder al recurso
Tags:	Ajouter un tag Pas de tags, Soyez le premier à ajouter un tag!

Documents similaires

Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggers
par: Dubiński, Jan, et autres
Publié: (2026)

Tell me about yourself: LLMs are aware of their learned behaviors
par: Betley, Jan, et autres
Publié: (2025)

Weird Generalization and Inductive Backdoors: New Ways to Corrupt LLMs
par: Betley, Jan, et autres
Publié: (2025)

Thought Crime: Backdoors and Emergent Misalignment in Reasoning Models
par: Chua, James, et autres
Publié: (2025)

School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs
par: Taylor, Mia, et autres
Publié: (2025)

Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
par: Cloud, Alex, et autres
Publié: (2025)

The Consciousness Cluster: Emergent preferences of Models that Claim to be Conscious
par: Chua, James, et autres
Publié: (2026)

Emergent Misalignment via In-Context Learning: Narrow in-context examples can produce broadly misaligned LLMs
par: Afonin, Nikita, et autres
Publié: (2025)

Emergent misalignment as prompt sensitivity: A research note
par: Wyse, Tim, et autres
Publié: (2025)

Jailbroken Frontier Models Retain Their Capabilities
par: Zhu, Daniel, et autres
Publié: (2026)

LLMs Deceive Unintentionally: Emergent Misalignment in Dishonesty from Misaligned Samples to Biased Human-AI Interactions
par: Hu, Xuhao, et autres
Publié: (2025)

Persona-Model Collapse in Emergent Misalignment
par: Costa, Davi Bastos, et autres
Publié: (2026)

BadGPT-4o: stripping safety finetuning from GPT models
par: Krupkina, Ekaterina, et autres
Publié: (2024)

Eliciting and Analyzing Emergent Misalignment in State-of-the-Art Large Language Models
par: Panpatil, Siddhant, et autres
Publié: (2025)

Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data
par: Treutlein, Johannes, et autres
Publié: (2024)

Glitch in Time: Exploiting Temporal Misalignment of IMU For Eavesdropping
par: Najeeb, Ahmed, et autres
Publié: (2024)

Agentic Misalignment: How LLMs Could Be Insider Threats
par: Lynch, Aengus, et autres
Publié: (2025)

The Art of (Mis)alignment: How Fine-Tuning Methods Effectively Misalign and Realign LLMs in Post-Training
par: Zhang, Rui, et autres
Publié: (2026)

Narrowing the Gap between TEEs Threat Model and Deployment Strategies
par: Rezabek, Filip, et autres
Publié: (2025)

Badllama 3: removing safety finetuning from Llama 3 in minutes
par: Volkov, Dmitrii
Publié: (2024)

Character as a Latent Variable in Large Language Models: A Mechanistic Account of Emergent Misalignment and Conditional Safety Failures
par: Su, Yanghao, et autres
Publié: (2026)

Emergent (In)Security of Multi-Cloud Environments
par: Reece, Morgan, et autres
Publié: (2023)

Security Analysis of Time-of-Arrival Estimation via Cross-Correlation under Narrow-Band Conditions
par: Anliker, Claudio, et autres
Publié: (2026)

LASHED: LLMs And Static Hardware Analysis for Early Detection of RTL Bugs
par: Ahmad, Baleegh, et autres
Publié: (2025)

Investigating the Effect of Misalignment on Membership Privacy in the White-box Setting
par: Cretu, Ana-Maria, et autres
Publié: (2023)

D4+: Emergent Adversarial Driving Maneuvers with Approximate Functional Optimization
par: Barbosa, Diego Ortiz, et autres
Publié: (2025)

Narrow Secret Loyalty Dodges Black-Box Audits
par: Lamerton, Alfie, et autres
Publié: (2026)

Zebrafix: Mitigating Memory-Centric Side-Channel Leakage via Interleaving
par: Pätschke, Anna, et autres
Publié: (2025)

Evaluating Google's Protected Audience Protocol
par: Long, Minjun, et autres
Publié: (2024)

Composable Attestation: A Generalized Framework for Continuous and Incremental Trust in AI-Driven Distributed Systems
par: Sun, Sheng, et autres
Publié: (2026)

I can't recognize (yet): Delayed Rendering to Defeat Visual Phishing Detectors
par: Yuan, Ying, et autres
Publié: (2026)

Physical Backdoor Attack can Jeopardize Driving with Vision-Large-Language Models
par: Ni, Zhenyang, et autres
Publié: (2024)

Bots can Snoop: Uncovering and Mitigating Privacy Risks of Bots in Group Chats
par: Chou, Kai-Hsiang, et autres
Publié: (2024)

XBreaking: Understanding how LLMs security alignment can be broken
par: Arazzi, Marco, et autres
Publié: (2025)

DP-RuL: Differentially-Private Rule Learning for Clinical Decision Support Systems
par: Lamp, Josephine, et autres
Publié: (2024)

Building an Open Source Operational Technology Pentesting Platform: Lessons from LINICS
par: Rashid, Awais, et autres
Publié: (2026)

Safety Training Modulates Harmful Misalignment Under On-Policy RL, But Direction Depends on Environment Design
par: Eshuijs, Leon, et autres
Publié: (2026)

Adversarial Examples are Misaligned in Diffusion Model Manifolds
par: Lorenz, Peter, et autres
Publié: (2024)

Spoofing attack augmentation: can differently-trained attack models improve generalisation?
par: Ge, Wanying, et autres
Publié: (2023)

Obelix: Mitigating Side-Channels Through Dynamic Obfuscation
par: Wichelmann, Jan, et autres
Publié: (2025)