OpenAI posts $445,000 safety role

OpenAI’s hiring pages show a company staffing for frontier-risk work in concrete, technical terms rather than broad policy language. Current job listings under its Preparedness and Safety Systems groups describe work on catastrophic misuse, self-improvement, misalignment, automated red teaming and internal agent threats, according to postings on OpenAI’s careers site. (openai.com) The specific $445,000 figure cited in secondary coverage was not visible in the OpenAI postings I could verify directly on Sunday, May 24, 2026. (openai.com) A currently accessible Preparedness listing for “Security Researcher, Agentic AI Threats” shows a compensation range of $293,000 to $405,000, while other Preparedness pages available through OpenAI’s site describe the scope of work without exposing a higher pay band in the text returned by search. (openai.com) ### So what is OpenAI actually hiring these people to do? OpenAI says its Preparedness team is part of the Safety Systems organization and is “tasked with identifying, tracking, and preparing for catastrophic risks related to frontier AI models.” A verified “Threat Modeler, Preparedness” posting says the role covers misuse domains including bio, cyber and attack planning, and also “loss of control, self-improvement, and other possible alignment risks from frontier AI systems.” (openai.com) A separate “Researcher, Frontier Cybersecurity Risks” posting says the team works on threats that “could scale to an extreme level of severity,” and describes measurement, mitigation and coordination as core functions. (openai.com) That listing says researchers would design and implement safeguards against model-enabled cyber misuse across OpenAI products. ### Why does “self-improvement” keep showing up in these postings? OpenAI’s own Preparedness Framework, updated on April 15, 2025, lists “AI Self-improvement capabilities” as one of its tracked categories for severe-harm risk. The company says it prioritizes capabilities when the risk is plausible, measurable, severe, net new, and either instantaneous or hard to remedy. (openai.com) The same framework also names future-facing research categories including autonomous replication and adaptation, undermining safeguards and long-range autonomy. That gives context for why a hiring post would ask staff to model risks tied to systems becoming more capable, harder to supervise or able to evade controls. (openai.com) ### Is this policy work, or hands-on engineering? OpenAI’s postings describe technical execution. The cyber-risk listing asks for deep learning experience and familiarity with model training and fine-tuning, while the agentic-threats role says candidates should be able to build prototypes and work across operating systems, cloud, containers and CI/CD. (openai.com) A current “Researcher, Automated Red Teaming” post is even more explicit. It says the job is to build scalable systems that uncover failure modes in models and safeguards, with an initial emphasis on jailbreak discovery, bio threat-development elicitation and evaluations tied to loss-of-control risks. (openai.com) ### What does the pay question tell us? The clearest verified number available from OpenAI’s site on Sunday was the $293,000-$405,000 range on the “Security Researcher, Agentic AI Threats” role in San Francisco. That figure places frontier-safety hiring in the same compensation conversation as senior engineering and research recruiting at major AI labs. (openai.com) OpenAI’s broader safety hiring also spans technical leads, data scientists and threat modelers, suggesting the company is staffing not just one headline role but a larger operating function around preparedness. The careers page remained live on May 24, 2026, and the Preparedness Framework remains the company’s public reference point for how those teams are organized. (openai.com 1) (openai.com 2) (openai.com 3)