OpenAI's Preparedness Team Dissolution Exposes the Governance Gap Anthropic's 133M Unfiltered Requests Made Real

The Day's Thesis

▶

Signal of the Day: Anthropic's bio-weapons filter was inactive for nearly a year, during which approximately 133 million unfiltered interactions reached its models — a quantified governance failure that arrives the same week OpenAI formally dissolved its Preparedness team.

The 30-Second Read:

• OpenAI disbanded its Preparedness team — the unit responsible for evaluating catastrophic model risks — reassigning staff and functions to existing groups
• Anthropic's bio-weapons and chemical-weapons filter was down for ~12 months, exposing 133 million requests from ~50,000 external contractors to unfiltered model outputs
• Gold closed Friday at $4,343.33/oz (+5.85% on the week) and Comex copper reached $6.59/lb (+1.42%), as BCM Resources reported 550m of copper mineralization grading 0.66% Cu in Utah
• Global EV sales hit 1.85 million units in July 2026 (+9% YoY), with European growth of 33% YoY masking a 27% North American contraction following the removal of U.S. federal tax credits

Two governance failures at the frontier AI labs — one structural, one operational — arrived simultaneously, providing the first concrete data point for what an AI safety vacuum looks like in practice. The magnitude of the Anthropic exposure (133 million requests, nearly a year) and the timing of OpenAI's reorganization are not coincident; they define the current risk perimeter of frontier model deployment.

---

AI & Research Frontier

The simultaneous dissolution of OpenAI's Preparedness team and Anthropic's 133-million-request filter failure define the outer boundary of enterprise AI governance risk as of mid-August 2026.

OpenAI's Preparedness unit — which ran structured red-teaming and catastrophic-risk evaluation against the company's own models — has been parceled out across existing groups, with multiple safety staffers departing. The reorganization leaves no dedicated team responsible for evaluating whether a frontier model could enable mass-casualty events; that function is now distributed, and internal sources describe a "burbling sense of responsibility and dread."