Orphan risks at the frontier of artificial intelligence: What diverging safety and compliance frameworks reveal about how AI companies choose the risks they prioritize

Companies prioritize risks they can measure and audit: The paper finds that frontier AI companies tend to focus on risks that are quantifiable and auditable in their self-authored safety documents, while less tractable risks like harmful manipulation are largely absent from these frameworks—even though these same companies articulate concerns about manipulation in compliance documents where law requires disclosure.

A "safety differential" exists between self-regulation and regulatory requirements: There is a significant gap between the risk landscape companies voluntarily select for themselves versus the risks regulators require them to address, suggesting that legal mandates—not internal prioritization—drive attention to certain harms.

Four filters determine which risks get managed: The paper identifies measurability, severity, auditability, and competitive cost as the primary factors determining which risks survive in company-authored frameworks, implying that risks failing these tests may be neglected regardless of their actual importance.

Multiple conflicting risk accounts from the same companies: Anthropic, OpenAI, Google DeepMind, and Meta maintain diverging accounts of what could go wrong with their technologies across different documents, revealing inconsistency in how these companies characterize their own systems' dangers.

Regulatory pressure, not corporate conscience, drives disclosure of difficult risks: Harmful manipulation and other complex risks appear in compliance documents but are conspicuously absent from voluntary safety frameworks, indicating that companies acknowledge these harms primarily when required by law rather than through independent risk assessment.