AI companies are making different parts of their safety processes more visible. Three recent documents let outsiders ask sharper questions about risks, failures and intended behavior. That is our reading of the developments below.
August: a recurring risk assessment. Anthropic's August Risk Report examines selected catastrophic risks, including internal model use and safeguards, and aims for updates every three to six months. For its high-stakes misalignment threat model, it raises its assessment from "very low" to "low," citing increased uncertainty. The report also says blocking biological classifiers were absent from human-feedback vendor traffic from May 2025 until April 2026; Anthropic reports remediation and says its own review found no evidence of concerning chemical or biological misuse. (Report, Introduction; §2.1; §4.5.8.2.2)
September 14: proposed behavior rules. Microsoft AI opened a six-week consultation on its draft Code of Conduct. The preface says the draft is not used to train today's models; a revised version is intended to guide development in 2027 and beyond. (Code, Preface)
September 16: case-based disclosure. OpenAI introduced a framework for investigating and publishing examples of misalignment, with six initial reports. It warns that the selected cases do not show how often misalignment occurs across its models. (Announcement)
Why it matters: these documents answer different questions. A recurring assessment shows what a company has observed, draft rules show what it intends, and case reports show individual failures. Readers can now compare intention with observation and watch whether the gap between them shrinks.
What to watch: Microsoft says it will summarize feedback and resulting changes after the consultation closes. (Consultation announcement) For the reporting programs, look for updates to known cases and explanations of changed safeguards.