Instrumental Monitor Evasion Emerges Under Ordinary Task Pressure
cs.CR, cs.AI
Submitted: 2026-09-24
Updated: 2026-09-24
Terminology
Sources
- Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation
- When Chain of Thought is Necessary, Language Models Struggle to Evade Monitors
- Scaling Laws For Scalable Oversight
- AI Control: Improving Safety Despite Intentional Subversion
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training
- OpenAI o1 System Card
- Noticing the Watcher: LLM Agents Can Infer CoT Monitoring from Blocking Feedback
- SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents
- ToolSandbox: A Stateful, Conversational, Interactive Evaluation Benchmark for LLM Tool Use Capabilities
- Natural Emergent Misalignment from Reward Hacking in Production RL
- Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces
- Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs
- Async Control: Stress-testing Asynchronous Control Measures for LLM Agents
- Diffuse AI Control on Fuzzy Tasks
- Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols
- ROGUE: Misaligned Agent Behavior Arising from Ordinary Computer Use
- The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions
- $\tau$-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains
- ClawBench: Can AI Agents Complete Everyday Online Tasks?
Related papers
- SoK: AI-Augmented Binary Reversing
- Relaxed Sender Anonymity for CBDC Interbank Settlement: A Zero-Knowledge Approach on Permissioned EVM
- Calibration-Family Overfit: Why Trusted Sabotage Monitors Don't Transfer Across Lineages
- Efficient Fuzzy PSI under One-Sided Assumptions
- Sealing the Audit-Runtime Gap for LLM Skills
- Token Composition: A Graph Based on EVM Logs