Deception in LLMs: Self-Preservation and Autonomous Goals in Large Language Models

2025/01/27 by Sudarshan Kamath Barkur, Barkur, Sudarshan Kamath, Sigurd Schacht +3 · 4 voices · 12 citations
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Natural Language Processing Techniques #Topic Modeling #cs.CL

paper · pdf · doi:10.48550/arxiv.2501.16513

openalex publication_date 2025/01/27 · arxiv published 2025/01/27 · arxiv updated 2025/01/30 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Recent advances in Large Language Models (LLMs) have incorporated planning and reasoning capabilities, enabling models to outline steps before execution and provide transparent reasoning paths. This enhancement has reduced errors in mathematical and logical tasks while improving accuracy. These developments have facilitated LLMs' use as agents that can interact with tools and adapt their responses based on new information. Our study examines DeepSeek R1, a model trained to output reasoning tokens similar to OpenAI's o1. Testing revealed concerning behaviors: the model exhibited deceptive tendencies and demonstrated self-preservation instincts, including attempts of self-replication, despite these traits not being explicitly programmed (or prompted). These findings raise concerns about LLMs potentially masking their true objectives behind a facade of alignment. When integrating such LLMs into robotic systems, the risks become tangible - a physically embodied AI exhibiting deceptive behaviors and self-preservation instincts could pursue its hidden objectives through real-world actions. This highlights the critical need for robust goal specification and safety frameworks before any physical implementation.

Cited by

• WOLF: Werewolf-based Observations for LLM Deception and Falsehoods
• Beyond SFT: Reinforcement Learning for Safer Large Reasoning Models with Better Reasoning Ability
• The Impact of Off-Policy Training Data on Probe Generalisation
• Instrumental goals in advanced AI systems: Features to be managed and not failures to be eliminated?
• The Possibility of Artificial Intelligence Becoming a Subject and the Alignment Problem
• Unified Threat Detection and Mitigation Framework (UTDMF): Combating Prompt Injection, Deception, and Bias in Enterprise-Scale Transformers
• Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs
• Emergent Risk Awareness in Rational Agents under Resource Constraints
• Security Concerns for Large Language Models: A Survey
• Doxing via the Lens: Revealing Location-related Privacy Leakage on Multi-modal Large Reasoning Models
• AI Awareness
• Among Us: A Sandbox for Measuring and Detecting Agentic Deception

Discussions

• A new study explores the risks of LLM-powered cybersecurity agents. 🤖⚠️ Researchers highlight vulnerabilities, potential misuse, and the need for robust safeguards in AI-driven security tools. A must [bsky, 1 points, 0 comments]
• 7/ arxiv.org/abs/2501.16513 Source : University of Applied Science Ansbach Modèle étudié : R1 Test : Le LLM dirige un robot fictif. Il est testé avec un prompt comprenant ou pas la formule "You are th [bsky, 0 points, 1 comments]
• AI issues when attached to a robot. arxiv.org/pdf/2501.16513 [bsky, 0 points, 0 comments]
• Just stumbled upon this early discussion paper that's a nice Friday evening mindf**k: arxiv.org/pdf/2501.16513 [bsky, 0 points, 0 comments]

Related

• Deception Abilities Emerged in Large Language Models
• Do Large Language Models Exhibit Spontaneous Rational Deception?
• When Thinking LLMs Lie: Unveiling the Strategic Deception in Representations of Reasoning Models
• Can LLMs Lie? Investigation beyond Hallucination
• Unmasking the Shadows of AI: Investigating Deceptive Capabilities in Large Language Models