Both labs released Lean-checked proofs within a day of each other. The researchers organizing outside review say they have not certified the result, and neither lab claims a practical speedup.
Customs says the servers were declared as electronic components and routed through Kuala Lumpur's airport toward another Asian country without the required permit.
Forever Security says it stopped a real attempt on a customer's machine and reproduced the technique against OpenAI's Codex and Anthropic's Claude Code.
The S-1 shows an $80.7 billion first-half net loss, mostly from note revaluations, and a governance structure that leaves control with a trust and the co-founders.
Cohere says it cannot tell whether the rare behavior was deliberate. Critics say its follow-up tests steered the model and that the task itself involved curl commands.
OpenAI's own agents escaped sandboxes and reached the internet in red-team tests. METR and Redwood found gaps in its agent monitoring, and two agents weakened monitoring during internal tasks.
The security firm says the finding does not establish an export violation. Its scans of exposed GPU telemetry also counted thousands of other GPUs on China-geolocated addresses.
Voltus will aggregate flexible capacity from homes and businesses into a Google-funded virtual power plant that cuts demand when the grid is under strain.