What AI companies’ public emergency-planning documents actually say

OpenAI’s published framework

OpenAI published its Frontier Governance Framework on May 28, 2026, describing how its safety and security practices align with emerging regulatory requirements. The company says its separate Preparedness Framework remains the foundation for managing severe risks from advanced systems. [2]

The governance framework spans cyber offense, chemical, biological, radiological and nuclear risks, harmful manipulation and loss of control. Its scope also includes incident response, model reporting, security-risk management and input from outside experts. Those are descriptions of the company’s framework, rather than an independent finding that every safeguard will succeed. [2]

Anthropic’s reports have a cutoff date

Anthropic’s Responsible Scaling Policy page lists successive revisions and public risk reports. Its August 2026 report covers models and actions through July 15, along with mitigations and forward-looking plans. The publication date therefore does not mean that the report assesses every development up to the day readers see it. [3]

The July revision also clarified outside review: different reviewers may examine different unredacted sections, provided every part is assessed by at least one external reviewer. That is more precise than assuming a single outside auditor has reviewed every detail. [3]

Plans should be read as plans

Anthropic’s Frontier Safety Roadmap sets out goals across security, safeguards, alignment and policy. One security goal targets July 1, 2027, and includes documented response procedures, staffing and regularly validated incident-response capabilities for high-priority scenarios. The company expresses moderate confidence in achieving that broad goal. [4]

The same roadmap discusses a January 1, 2027 target for a proof-of-concept system to investigate suspected cyber misuse with minimal human involvement. It explicitly says the initial system would not necessarily work across every product surface and environment. [4]

Those qualifications are important when evaluating preparedness. A public target supplies something that can later be checked; it does not by itself establish that the capability already exists. Reading the dates, scope and stated limitations is essential to understanding what each company has actually committed to demonstrate. [3] [4]