Can AI Agents Be Trusted to Follow the Rules?
Five new AI governance papers raising uncomfortable questions | Edition #346

TL;DR
- ScrapeBench is the first agentic benchmark to evaluate AI agents' legal compliance with website scraping policies.
- A systematic review identifies 95 existential risk scenarios from AI, grouping them into eight families.
- Studies suggest AI models have varying attitudes towards the law and are highly susceptible to conditioning.
- Authorization drift in LLM agents is linked to legal provisions, highlighting the need for trajectory-level training and runtime authorization controls.
- A participatory study reveals tensions in translating the EU AI Act into technical requirements, human oversight, and the validity of compliance evidence.