The Superalignment Team Is Gone. Its Replacement Reports to the Board.

Regulatory Pressure Forces OpenAI's Hand

On August 29, 2024, the U.S. AI Safety Institute at NIST announced memoranda of understanding with OpenAI and Anthropic, the first formal agreements giving a federal body pre-release access to frontier models. Each MOU requires the companies to share major new models before and after public deployment, enabling collaborative evaluation of capabilities, safety risks, and mitigation methods. "An important milestone as we work to help responsibly steward the future of AI," said institute director Elizabeth Kelly.

The move came 270 days after President Biden's executive order on AI, which tasked NIST with developing testing standards for dual-use foundation models. In July, the agency released draft guidance (NIST AI 800-1) outlining seven voluntary practices for preventing misuse, covering biological weapons, cyber operations, and nonconsensual intimate imagery. The guidance explicitly targets developers of "dual-use foundation models," a category that includes OpenAI's GPT series.

The EU AI Act entered force after publication in the Official Journal on July 12, 2024, starting a 24-month countdown to full enforcement. The regulation imposes tiered obligations on general-purpose AI models based on compute thresholds, with full roll-out foreseen by August 2027. California's SB 1047, passed by the legislature August 28 and awaiting Governor Newsom's decision, would add state-level safety testing mandates for models trained above a defined compute budget.

OpenAI responded with public statements. Sam Altman posted on X: "We are happy to have reached an agreement with the US AI Safety Institute for pre-release testing of our future models." Chief Strategy Officer Jason Kwon told CNBC the company "strongly supports the U.S. AI Safety Institute's mission." But voluntary cooperation and public endorsements do not satisfy a regulatory regime that now expects documented, repeatable oversight processes embedded in every release cycle. That requirement forced the build-out of a dedicated safety-governance workforce — distinct from research — capable of producing the evidence trails, model cards, and cross-functional sign-offs that regulators and the institute will demand.

The Strategic Hire: Technical Program Manager, Safety Governance

OpenAI's Safety Systems team has posted at least three distinct program-management roles, each targeting a different layer of the safety-to-production pipeline. The most governance-focused listing (titled Technical Program Manager, Governance on the Khosla Ventures board and Technical Program Manager, Safety Systems Engineering on OpenAI's own careers page) makes the mandate explicit: "streamline our full safety governance process and integration of various safety research and mitigations into our ChatGPT, API, and any frontier models."

The role sits inside Safety Systems, the group responsible for evaluations, safeguards, red-teaming, and deployment decisions. Its core deliverables read like a compliance officer's checklist: track safety research progress and risk tables; oversee human-data campaign quality; standardize the risk-assessment lifecycle from pre-launch review through post-launch follow-up; manage pre-launch safety reviews; share launch calendars and key evaluations with Microsoft; and produce documentation thorough enough to publish safety learnings, standards, datasets, and benchmarks publicly.

Qualifications signal the hybrid profile OpenAI wants: an advanced degree in a hard science (PhD "advantageous") or equivalent engineering track record; proven delivery of high-profile, complex technical projects on tight deadlines; fluency partnering with fundamental research teams; and "strong knowledge of content integrity and moderation, including industry best practices and regulatory guidelines."

A companion Program Manager, Safety role narrows the aperture to "streamline our safety review process," coordinating cross-functional launch readiness across model and product releases, building operating rhythms so past-launch learnings feed back into automated safeguards and tooling, and designing incident-management and mitigation-update programs. Together, the postings reveal a deliberate split: one TPM owns the governance infrastructure (risk tables, compute roadmaps, partner coordination, public artifacts); the other owns the review operating rhythm (launch gates, stakeholder synthesis, continuous process improvement).

Both roles are based in San Francisco on a three-day office hybrid. Both list "care about AGI safety" as a cultural filter. And both treat regulatory guidelines (not just research best practices) as required knowledge.

Who Builds the Oversight Layer?

OpenAI's Safety Oversight Research team sits inside Safety Systems with a mandate the job posting states plainly: "fundamentally advance our capabilities to maintain oversight over frontier AI models, and leverage these advances to ensure OpenAI's deployed models are safe and beneficial." The team targets four research fronts (human-AI collaboration, reasoning, robustness, and scalable oversight) to keep pace with model capabilities.

The Researcher, Safety Oversight role is a senior position priced at $295K–$445K plus equity. The posting asks for a Ph.D. or equivalent in computer science or machine learning, four-plus years in AI safety (specifically RLHF, human-AI collaboration, fairness and bias), and four-plus years of research engineering with Python. This is not an entry-level safety hire; it is a builder of the oversight infrastructure itself.

Day-to-day work centers on five pillars. Develop and refine AI monitor models that detect known and emerging patterns of misuse and misalignment. Set research directions that make systems safer, more aligned, and more robust. Design red-teaming pipelines that stress-test safety systems end to end. Improve models' ability to reason about human values and apply those improvements to practical safety challenges. Coordinate across Trust & Safety, legal, policy, and other research teams so products meet the highest safety standards before launch.

The language signals a shift from publishing papers to shipping guardrails. "Monitor models" and "red-teaming pipelines" are production-grade tooling. "Reason about human values" moves alignment from theory into evaluatable model behavior. The cross-functional requirement (T&S, legal, policy) means this researcher operates at the seam where research becomes enforceable product policy.

Governance Differs from Research

OpenAI's safety update from May 2024 lists "alignment and safety research" as its second practice, distinct from the "systematic approach for safety" that implements measures "at every stage of the model's life cycle, from pre-training to deployment." That distinction is the fault line between the old workforce and the new one.

Safety research at OpenAI has historically meant the Superalignment team (formed 2023, disbanded 2024) and the Mission Alignment team (formed September 2024, disbanded February 2026). Their output was technical: robustness to jailbreaks, refinement of RLHF and DPO techniques, red-teaming protocols, the Preparedness Framework's risk thresholds. The research question was does the model fail?, measured in error rates, refusal patterns, capability evaluations. More than 70 external experts assessed GPT-4o through red-teaming; the results fed back into evaluation design.

Safety governance asks different questions. "Who authorized this system for this use. What risks were accepted and by whom. How tradeoffs were evaluated. What evidence exists to justify continued operation," wrote Sneh Lata on LinkedIn. The Board-level Safety and Security Committee, created in May 2024 and chaired by Sam Altman, Bret Taylor, Adam D'Angelo, and Nicole Seligman, exists to answer those questions. Its first 90-day mandate was to evaluate development processes and safeguards for the next frontier model — not the model's behavior, but the organization's behavior around it.

The cross-functional Safety Advisory Group, formalized in the Preparedness Framework, sits between the two. It reviews capability reports and makes deployment recommendations. Company leadership decides; the Board oversees. That chain — technical review → leadership decision → Board oversight — is governance infrastructure. It did not exist when the Superalignment team was writing papers on scalable oversight.

The new roles make the split operational. The Technical Program Manager, Governance is tasked with integrating safety research into ChatGPT, API, and frontier models, taking finished research and embedding it in release gates, monitoring pipelines, and incident response. The Researcher, Safety Oversight role focuses on maintaining oversight over frontier models, identifying and mitigating misuse and misalignment — not discovering new alignment techniques, but designing the organizational controls that determine when and how models ship.

OpenAI's own language signals the shift. The safety update emphasizes "practical alignment, safety systems, and post-training research" (research that feeds product). The Frontier Governance Framework maps safety practices to EU AI Act and California regulatory requirements. That is compliance engineering, not model science.

The Mission Alignment team's disbanding confirms the trajectory. Its six or seven members were reassigned "to different parts of the company" doing "similar work"; but the function moved from a dedicated team to embedded responsibility. An analysis of the restructuring characterized it as a shift from "Superalignment philosophy" to "Safety-by-Design," where protocols live inside core engineering teams building GPT-4o and the o3 series. The Chief Futurist role (Josh Achiam's new title) handles long-term forecasting; the day-to-day governance now runs through product and operations.

For engineers, the difference shows up in workflow. Safety research produces model cards, system cards, evaluation suites. Safety governance produces approval gates, escalation paths, risk-acceptance records, Board reporting packages. One is peer-reviewed; the other is audited.

Reporting Lines and Cross-Functional Integration

The safety governance roles are not parked in a standalone compliance silo. OpenAI's updated structure, finalized in October 2025, places the Safety and Security Committee (SSC) as a standing committee of the OpenAI Foundation board, chaired by Dr. Zico Kolter, with explicit governance authority over "the safety and security practices of all of OpenAI, including OpenAI Group." That gives the SSC a direct line to the nonprofit board, not to a VP of engineering.

Operationally, the cross-functional execution sits with the Safety Advisory Group, which OpenAI describes as the body that "reviews model capability reports and makes recommendations ahead of deployment" under its Preparedness Framework. The group spans infrastructure, applied engineering, legal, policy, and product, a structure the company itself calls "cross-cutting." Technical Program Managers in security and safety are explicitly tasked with "aligning teams and delivering execution at scale" across those same functions.

Chris Clark's recent appointment as Head of Nonprofit and Strategic Initiatives adds another signal: he leads "the operations of OpenAI's nonprofit parent and key strategic projects including our relationships with mission-aligned partners." That role bridges the Foundation's governance mandate (where the SSC lives) with the Group's product cadence. Meanwhile, Brad Lightcap as COO oversees "operational functions and ensuring that strategic initiatives are executed efficiently," per the org chart, putting the COO's office in the path of any governance-to-production workflow.

In practice, a safety governance hire (whether a TPM for Governance or a Researcher for Safety Oversight) will operate at the intersection of the SSC's board-level mandate, the Safety Advisory Group's pre-deployment reviews, and the product/engineering leads (CTO Mira Murati, VP Product Peter Welinder) who own launch decisions. The reporting line is not a single manager; it is a matrix anchored at the Foundation board, mediated through Strategic Initiatives, and executed cross-functionally.

How Peers Are Staffing for Compliance

Anthropic's careers page lists a dedicated "Safeguards (Trust & Safety)" category with 15 open roles: Biological Safety Research Scientist, Technical CBRN-E Threat Investigator, Technical Cyber Threat Investigator, and a Technical Program Manager for Safeguards infrastructure and evals. The company's Legal team carries an AI Compliance Officer in Dublin and a Compliance Governance & Oversight Lead spanning San Francisco, New York, and Washington, DC. Zero G Talent's board shows 327 Anthropic roles with a median salary of $405k; 28 were added in the past week.

Zero G Talent's data shows Mistral's board presence is smaller: one role listed with a $62k–$125k band, and five new postings in the last seven days including that Deputy Director slot.

The Future of Life Institute's 2024 AI Safety Index evaluated six leading general-purpose AI companies across six governance domains. The World Economic Forum flags a growing demand for trained professionals to address AI governance needs. Stanford's 2024 AI Index Report arrives at a moment when AI's societal influence has never been more pronounced. These signals align: safety is moving from research into operational compliance.

OpenAI's new safety-governance hires (the TPM for Governance and the Researcher for Safety Oversight) mirror the structure Anthropic has already institutionalized. The sector is converging on a model where safety governance reports into operations, not just research, and where regulatory-grade roles carry engineering titles.

The memoranda of understanding announced in August were the signal. The hiring — program managers who write launch gates, researchers who build monitor models, committees that answer to a nonprofit board — is the response. Safety governance has become a permanent operational function, staffed by engineers with regulatory-grade mandates, and the workforce is being built in plain sight.

Working in AI? Zero G Talent tracks the openings: browse AI jobs, openings at OpenAI, Anthropic and Mistral AI, and the people building the field.