OpenAI Hiring Researchers to Curb Autonomous AI Self-Improvement
- •OpenAI is hiring a Preparedness researcher with a salary of up to $445,000 to mitigate risks from self-improving AI systems.
- •The role involves tracking internal AI automation and defending against data poisoning, supporting goals for a fully autonomous researcher by 2028.
- •Industry focus on recursive self-improvement grows as METR reports that frontier AI task complexity doubles every seven months.
OpenAI has posted a job listing for a researcher within its Preparedness safety team, offering a salary range of $295,000 to $445,000 (approximately ₹2.5 crore to ₹3.7 crore). The role focuses on recursive self-improvement, a concept where AI systems research, design, and train superior versions of themselves without significant human intervention. The company describes this as urgent work, noting that staff must reason about future risks that may not yet exist.
The initiative aligns with OpenAI's stated research goals. In October, CEO Sam Altman announced an objective to deploy an automated AI research intern across hundreds of thousands of chips by September 2026, with a target of developing a fully autonomous AI researcher by March 2028. This move builds on the commercial success of the company's Codex coding tools, which currently serve as a major revenue stream. The new researcher is expected to track the automation of technical staff, monitor the usage of AI coding tools internally, and defend models against data poisoning, which involves corrupting training datasets.
Industry attention on autonomous development has intensified as frontier AI capabilities accelerate. Research from METR, a laboratory specializing in AI capabilities, indicates that the duration of tasks frontier models can complete doubles approximately every seven months. Demis Hassabis, CEO of Google DeepMind, recently stated that humanity is approaching the "foothills of the singularity," where AI improvement cycles exceed human speed. Other organizations are also exploring these risks, with Anthropic publishing research in April on using AI to oversee more powerful models. Jack Clark, a co-founder at Anthropic, estimates a 60 per cent probability that AI-led research and development will occur without human oversight by the end of 2028.