Alex Mallen argues in an October 2 post on Redwood Research that additional safety options do not automatically make AI safer. What matters is which options developers actually choose.
His model compares safety and usefulness. Capability improvements could enable stronger safeguards while also increasing the incentive to accept more risk for more useful systems.
With credible restraint on further capability growth, some research could instead improve safety. Mallen presents this as a conditional future scenario, not a justification for today's race.
This risk signal records a research assessment, not a new incident. Mallen acknowledges imperfect information and the difficulty of measuring safety. The model provides no empirical probability of losing control.