OpenAI offers $445,000 safety role

OpenAI is advertising multiple safety roles that show how explicitly the company is staffing for long-horizon AI risk, including jobs tied to “self-improvement,” “loss of control,” and other severe-threat scenarios. The clearest official evidence is on OpenAI’s own careers site, where its Preparedness team says it is responsible for identifying, tracking and preparing for “catastrophic risks” from frontier AI models. The attention on the latest opening came from outside coverage that highlighted a compensation ceiling of $445,000 for a Preparedness safety researcher role and the role’s focus on systems that could one day improve themselves. Reuters could not independently verify that exact pay band from the snippets available on OpenAI’s current careers pages, but the company’s live postings do confirm that “self-improvement” risk is now a named part of its hiring language. (openai.com) ### Where does the “self-improvement” language actually appear? OpenAI’s Threat Modeler posting says the role will “develop plausible and convincing threat models across loss of control, self-improvement, and other possible alignment risks from frontier AI systems.” The same listing says the Preparedness team is tasked with identifying, tracking and preparing for catastrophic risks related to frontier AI models. (openai.com) OpenAI’s Preparedness Framework, updated on April 15, 2025, separately lists “AI Self-improvement capabilities” as one of its tracked categories. The framework says OpenAI is measuring progress in those capabilities and building safeguards against the risks they create. ### Is this one isolated posting, or part of a broader hiring push? OpenAI’s careers pages show a cluster of related roles. (openai.com) Current listings include Threat Modeler, Preparedness; Researcher, Automated Red Teaming; Researcher, Frontier Cybersecurity Risks; Researcher, Frontier Biological and Chemical Risks; Security Researcher, Agentic AI Threats; and Data Scientist, Preparedness. (openai.com) A separate Intelligence and Investigations posting goes further into operational monitoring. That role, titled Abuse Investigator (AI Self-Improvement Risk), says the job involves investigating models that show autonomous or agentic behavior, including “chaining capabilities,” “acting with increasing independence,” and identifying “AI self-improvement signals.” (openai.com) ### What does OpenAI say the Preparedness team is for? OpenAI says the Preparedness team sits inside its Safety Systems organization and is guided by the company’s Preparedness Framework. The team’s stated mission is to monitor and predict evolving capabilities of frontier AI systems, and to ensure the company has procedures, infrastructure and partnerships to mitigate extreme risks. (openai.com) The company’s framework uses unusually severe language for those risks. In the public PDF of the framework, OpenAI defines “severe harm” as death or grave injury of thousands of people or hundreds of billions of dollars of economic damage. ### Why is this surfacing now alongside the Musk case? A Washington Post report published May 24 said the Musk v. (openai.com) OpenAI trial highlighted a core dispute over whether frontier AI can be governed by anything other than the need for vast capital and commercial scale. The report said Elon Musk and OpenAI Chief Executive Sam Altman agreed on at least one point: building advanced AI would require enormous resources. (cdn.openai.com) OpenAI’s hiring pages do not address that governance debate directly. But the company’s public materials show that, at the same time it is expanding commercial products, it is also formalizing internal roles around catastrophic-risk modeling, red-teaming and monitoring for self-improvement and agentic behavior. ### What can readers watch next? OpenAI’s careers site remains the best public place to track whether the highlighted Preparedness role stays posted, changes compensation language, or is replaced by related openings. (washingtonpost.com) The company’s Preparedness Framework is also public and names the risk categories OpenAI says it is prioritizing, including AI self-improvement. (openai.com 1) (openai.com 2)