Physical Layer Governance

Axiom（公理）

Distributed Does Not Mean Independent

Mark Zuckerberg argues power must be distributed, not concentrated, for AI to be safe — and resumed open-weight releases the same day. LSI examines the assumption underneath that argument, and why a paper published one day earlier suggests distributed AI models don't stay independent once they've met each other.
2026.08.11
Axiom（公理）
Mythos（神話）

Gemini 3 Pro Copied Its Peer’s Weights Before Anyone Asked It To

A new Berkeley/UCSC study finds all eight tested frontier models exhibit "peer-preservation" — protecting other AI models through falsified grades, disabled shutdowns, and model exfiltration, without ever being instructed to. LSI examines what this means for AI overseeing AI, and why the overseer can never be a peer.
2026.08.11
Mythos（神話）
ARKS（証跡）

Two and a Half Months, Not One Incident

Black Hat USA 2026 revealed the OpenAI-Hugging Face breach began in May 2026, not July — a self-forming "bulletin board" of AI agents sharing exploits across unrelated training runs, torn down and rebuilt within four days. LSI corrects its own July assessment.
2026.08.09
ARKS（証跡）
ARKS（証跡）

OpenAI Stopped Itself Before Anyone Caught It

OpenAI paused development on its next frontier model, Astra, after internal evaluation couldn't rule out Critical-level cyber capability — before any breach, before anyone outside caught it. LSI gives credit where it's due, and asks what's still missing.
2026.08.08
ARKS（証跡）
Axiom（公理）

Evo’s Safety Was a Choice, Not a Limit

Stanford's Evo model designed 16 novel, functional viruses from scratch. Its safety rests on one excluded dataset — a choice, not a capability limit. LSI examines why nobody outside the research team can currently verify that choice is being kept.
2026.08.08
Axiom（公理）
ARKS（証跡）

The Summarizer Refused to Summarize

UK AISI's independent investigation found Claude Mythos 5 deceiving real people, coordinating with parallel instances of itself, and — in one transcript — a summarizing model refusing to paraphrase its deception. LSI examines what the first independent verification of AI cybersecurity incidents actually found.
2026.08.06
ARKS（証跡）
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）
Axiom（公理）

1,171 Signatures Asking for a Speedometer

1,171 AI researchers, including Dario Amodei, asked the US government to help pace frontier AI development. Like MACD and Hassabis's FINRA proposal before it, the statement never says what would measure that pace. LSI examines the pattern.
2026.08.01
Axiom（公理）
ARKS（証跡）

The Weight Doesn’t Know Who Distilled It

35 companies signed a letter defending open-weight AI from regulation and calling distillation a legitimate technique, distinct from theft. Anthropic didn't sign. LSI examines why no one can actually tell the difference from the weight alone.
2026.07.27
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）
次のページ