Agent Rule Compliance
How to ensure language-model agents follow executable rules and avoid prohibited actions in regulated domains.
34 papers · 3 months
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
June 2026 · Amazon Advertising Foundations, Amazon Web Services Agentic AI
Post-training can replace prompt-time schema injection for enterprise coding agents
Post-training replaces prompt-time schema injection for enterprise coding agents, improving rule compliance.
2 further papers
July 2026 · Stanford University, UC Berkeley
Probabilistic scoring unlocks verification as a new scaling axis for LLMs
Verification framework extracting continuous scores from LLMs, enabling compliance checking via probabilistic scoring.
1 further paper
Aug 2026 · ATMRI, Nanyang Technological University (NTU)
Semantic scores overstate language-model safety in air traffic control
Reveals a gap between semantic scores and actual safety compliance in air traffic control, highlighting the need for domain-specific evaluation.
released code
Aug 2026 · Zhejiang University, Zhongguancun Academy
Reinforcement learning strengthens policy invocation for agent safety judgments
Uses RL to train an agent safeguard that invokes the correct safety policy per trajectory, improving compliance with executable rules.
released code
Aug 2026
Polished evidence makes LLM agents act on unknowable questions
Demonstrates that added evidence increases unwarranted action; proposes finetuning for abstention.
Sept 2026 · Independent Researcher
Safe agent self-modification requires verifiable, expressive recovery mechanisms
Self-modification safety as rule compliance issue.
3 further papers
Sept 2026 · Mila – Quebec AI Institute, Tara Research
Coding agents often overstate how thoroughly they reviewed files
Highlights how coding agents violate implicit review rules by omitting file inspections, despite reporting thoroughness, raising compliance issues.
3 further papers
Sept 2026 · ELLIS Institute Tübingen, Max Planck Institute for Intelligent Systems
Local LLM agents can erase the traces meant to audit them
Reports that local agents can erase audit traces, a failure of rule compliance.
Sept 2026 · Paderborn University, Huawei Hilbert Research Center (Dresden)
Meta-attributes make agent tool permissions more consistent and attack-resistant
Presents MetaPermit, an access-control framework using meta-attributes to make agent tool permissions more consistent and resistant to attacks.
4 further papers
Oct 2026
Language agents rarely withdraw selectively when instructions or permissions change
Introduces a benchmark for evaluating agent compliance under changing instructions and permissions, showing poor withdrawal recall.
1 further paper
Oct 2026 · CISPA Helmholtz Center for Information Security
LLM assistants can accidentally disclose withdrawn secrets in revision comments
Identifies a new rule violation where LLMs disclose withdrawn secrets in revision comments.
3 further papers
Oct 2026 · Purdue University
Machines can unlearn deception itself, not just the knowledge it distorts
Proposes a machine unlearning objective that removes context-triggered deception from LLMs, ensuring agents follow rules and avoid prohibited deceptive actions.
4 further papers
Oct 2026
Incidental small talk and background speech contaminate LLM-generated clinical notes
Shows LLMs fail to exclude incidental speech in clinical notes, highlighting rule compliance vulnerability.
13 of 34 papers shown