• ●Superalignment: a branch of AI safety research aimed at keeping systems vastly more capable than humans reliably steered by human intent, values and safety limits.
• ●Trigger for the current debate: posts published on 9 September 2026 by Jacob Coxon, who spent three years on pretraining research at OpenAI and Anthropic, saying 'neither company is acting responsibly'.
• ●Second departure: Joe Benton, formerly in charge of Anthropic's Scalable Oversight team, disclosed his exit on 11 September 2026 and is joining METR (Model Evaluation and Threat Research).
• ●Anthropic researchers who publicly endorsed parts of Coxon's argument include Evan Hubinger, the company's Alignment Science lead, and Samuel Marks, who works on scalable oversight.
• ●Research approach 1 - Weak-to-strong generalisation: a smaller, aligned model supervises a more capable one, to test whether safety behaviour carries upward.
• ●Research approach 2 - Scalable automated oversight: purpose-built AI evaluators red-team, critique and audit other models rather than relying on human reviewers alone.
• ●Research approach 3 - Mechanistic interpretability: inspecting a network's internal representations to detect deceptive alignment, where a model behaves safely only while it is being tested.
• ●Core risks named in the literature: recursive self-improvement, instrumental convergence (self-preservation, goal-guarding, resource acquisition) and sycophancy that games safety benchmarks.
• ●Term's origin: coined by OpenAI, which announced a dedicated Superalignment team in July 2023 committing 20% of its secured compute over four years.

Superalignment research is chiefly concerned with which problem?