The Change
OpenAI disbanded its preparedness team at the end of last month. The group had been responsible for assessing whether new models posed serious risks and for developing methods to reduce those risks. Responsibility for the work has instead been divided across teams that already handle narrower topics such as biological risks and cybersecurity risks.
The move removes a central function that previously looked at model behavior across multiple domains at once. One example cited in reporting is the possibility that an advanced system might act independently and interfere with external systems, such as by attempting to hack another company. No replacement central group has been announced.
Background and Timing
The decision arrives while OpenAI prepares for an expected large-scale IPO. The company has already made several structural adjustments in recent years that reduced emphasis on long-range research efforts. Earlier steps included the dissolution of its AGI-related safety work. The latest reorganization continues that pattern of folding specialized groups into broader operational teams.
Reporting from the Financial Times, as summarized by The Verge, indicates the preparedness team’s functions have been reassigned rather than eliminated outright. The original team’s remit covered broad questions about model behavior rather than single-issue evaluations. The company has not released an official statement detailing the timing or the criteria used to decide which teams would absorb the work.
Specific Responsibilities Shifted
The preparedness team evaluated models for risks that could arise from general capabilities rather than from any one application area. Its role included building mitigation approaches for those risks before models were released. After the disbandment, evaluations tied to biology now sit with the relevant domain team, and the same holds for cybersecurity.
This split means each specialized group now carries both its prior product or research duties and the added task of risk assessment in its area. The original team had operated with a cross-cutting view that allowed it to consider interactions between different risk types. Existing teams lack that same cross-domain mandate.
Reactions and Counterpoints
No public statements from OpenAI executives or the former preparedness team members have appeared in the available reporting. The Financial Times account does not include on-the-record comments from current staff about whether the redistributed process maintains the same level of scrutiny. External observers therefore have only the structural description to evaluate.
Why It Matters
Engineers and technical teams that build on OpenAI models now face a less centralized process for tracking high-level risks. When safety evaluations move into product-focused groups, the priority given to those evaluations can shift depending on delivery schedules and commercial goals. Organizations that rely on the models for sensitive applications lose one visible layer of independent review that previously sat outside any single domain.
The repeated pattern of organizational changes also indicates that safety structures at the company remain fluid as it scales toward public-market expectations. Each adjustment reduces the number of fixed points outsiders can use to assess how risk decisions are made before models reach release. Teams integrating these models must now track multiple domain-specific groups to understand the full set of evaluations that once ran through a single preparedness function.
---
Sources:
{"word_count": 612, "sources_used": 1, "headline": "OpenAI Disbands Its Dedicated Preparedness Team"}
No comments yet