A readable archive
AI System Cards
AI companies publish their safety work as long PDFs — a system card when a model ships, a risk report in between — and reading a long PDF kind of sucks. This archive re-typesets each one as a faithful, responsive web page — every sentence, table, figure, and footnote, with deep links back to the source — plus clean markdown for both humans and machines (llms.txt).
Fair warning: these are auto-converted, so there’s almost certainly the odd artifact or transcription slip in here. They came out pretty damn good all the same — and the original PDF is one click away on every page.
-
Risk Report: August 2026
Anthropic’s Risk Report for August 2026 (public redacted edition), published under version 3.4 of the Responsible Scaling Policy. Unlike a system card, it assesses catastrophic risk from Anthropic’s activities as a whole — misalignment in high-stakes settings, automated AI R&D, and chemical and biological weapons production — across all models including Claude Mythos 5 and the unreleased internal “Model 2”, together with security controls, deployment safeguards, and cross-cutting considerations.
-
Claude Opus 5
System card for Claude Opus 5, an upgrade to Claude Opus 4.8 from Anthropic. Covers RSP evaluations, cyber capabilities, safeguards and harmlessness, agentic safety, alignment assessment, model welfare, and capabilities.
-
Claude Fable 5 & Claude Mythos 5
System card for Claude Mythos 5 and Claude Fable 5, two configurations of a new large language model from Anthropic. Covers RSP evaluations, cyber capabilities, safeguards and harmlessness, agentic safety, alignment assessment, model welfare, and capabilities.