Startup Commercializes Guardrail-Stripped Frontier Models as a Hosted Service

Skynet Chance (+0.09%): Commercializing the removal of safety alignment from near-frontier open-weight models shows that current guardrails are a thin, easily stripped layer rather than deep alignment, and reduces the friction for misuse to a credit card and a browser. It normalizes the deployment of deliberately unaligned systems at scale with essentially no access controls.

Skynet Date (-1 days): Turning an underground technique into a low-friction commercial service, with venture funding in talks, accelerates the arrival of widely available unaligned models that assist with cyber and bio harms. That pulls forward the window in which serious AI-enabled harm becomes routine.

AGI Progress (0%): Abliteration strips refusal behavior rather than adding capability, and practitioners quoted note it can degrade a model's knowledge and performance, so it contributes almost nothing to AGI-relevant capability. The implicit signal that open-weight GLM-5.3 is near-frontier is the only mildly relevant datapoint.

AGI Date (+0 days): Unrestricted access to strong open weights may marginally speed agentic experimentation and tooling, but the news involves no new algorithm, compute scale-up, or training advance. The effect on the AGI timeline is close to negligible.