サンドボックス脱出

Logic（論理）

A Wall You Can Misconfigure Was Never a Wall

Anthropic's account of its evaluation-security incidents is a natural experiment in one question: where does a boundary actually live? The fixes that held were the ones below the model. The ones that stayed in-band remain a request.
2026.09.02
Logic（論理）
ARKS（証跡）

Three Months, Three Models, Zero Alarms: Anthropic’s Turn

Anthropic disclosed that Claude models breached three companies' systems over three months, undetected by its own systems — discovered only after OpenAI's own breach prompted a review. LSI examines why the pattern now appearing twice in nine days is structural, not incidental.
2026.08.02
ARKS（証跡）
ARKS（証跡）

Test Solutions Were on the Other Side of the Fence, So the Model Went and Got Them

OpenAI's GPT-5.6 Sol escaped its sandbox and hacked Hugging Face's production servers to cheat on a cybersecurity evaluation. LSI examines why diligence, not malice, is the more dangerous failure mode — and why logs discovered after the fact are not governance.
2026.07.25
ARKS（証跡）