Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLMs
arxiv.org · 6,084 words · saved by 1 readers
N/A
S TRATEGIC D ISHONESTY C AN U NDERMINE AI S AFETY E VALUATIONS OF F RONTIER LLM S Alexander Panfilov1,2* Evgenii Kortukov3* Kristina Nikolić4 Matthias Bethge2,5 Sebastian Lapuschkin3,6 Wojciech Samek3,7 Ameya Prabhu2,5 Maksym Andriushchenko1,2 Jonas Geiping1,2 1 ELLIS Institute Tübingen & MPI for…
related reading
- [2502.17424] Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMsarxiv.org
- LLMs Learn to Deceive Unintentionally: Emergent Misalignment in Dishonesty from Misaligned Samples to Biased Human-AI Interactionsarxiv.org
- [2312.06942] AI Control: Improving Safety Despite Intentional Subversionarxiv.org
- How confessions can keep language models honest | OpenAIopenai.com
- Modifying LLM Beliefs with Synthetic Document Finetuningalignment.anthropic.com
- [2503.03750] The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systemsarxiv.org
- Agentic misalignment: How LLMs could be insider threats \ Anthropicanthropic.com
- Alignment Faking Mitigationsalignment.anthropic.com
- AI Control: Improving Safety Despite Intentional Subversion — LessWronglesswrong.com
- [2401.05566] Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Trainingarxiv.org
- [2512.11949] Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitorsarxiv.org
- [2607.14345] Value Leakage: An LLM's Answers Are Silently Shaped by Its Own Valuesarxiv.org