Fetching from the wire…
Research2026-07-29 · source-backed
Anthropic's Frontier Red Team published a previously unknown attack on a NIST post-quantum signature candidate, found by a multi-agent system with a human collaborator who is not a lattice cryptography specialist. Claude also produced "Möbius Bridge," a fingerprinting technique making an existing meet-in-the-middle attack on 7-round AES-128 between 200 and 800 times faster, at roughly a billion output tokens over three days and about $100,000 per primary attack. Neither result touches production: HAWK is undeployed, the AES result is reduced-round. The Hacker News reported it could not find any independent reproduction at review time and the disclosure thread had no replies. That caveat belongs in every summary of this.
Each link below shares sources, entities, or timing with this story.
Buried in the writeup: Claude worked almost entirely autonomously after researchers built a scaffold, generating several hundred million tokens over three days from three substantive human messages, ultimately reaching one billion output tokens. The attack came out 200 to 800...
Anthropic researchers used Claude Mythos to discover cryptographic weaknesses in the HAWK signature scheme and deliberately weakened AES variants. Human intervention was minimal and mostly consisted of encouraging the model not to give up and to push for publishable quality, n...
Green's July 29 assessment splits the two results sharply. He calls the HAWK attack significant precisely because HAWK was advancing toward standardization and "none of the ingredients are exotic," meaning the model won by thoroughly applying existing tools rather than inventi...
The letter to Senators Tim Scott and Elizabeth Warren, dated June 10 and surfacing publicly this week, frames it as model distillation run against Claude at scale (Anthropic). A related claim pegs it at 28.8 million fraudulent exchanges, though that figure is single-sourced an...
$688 billion in hyperscaler AI capex against $110 billion of revenue. No capex breakeven until 2031 or 2032. Those are Scale Venture Partners' Rory O'Driscoll's numbers from SaaStr AI 2026, and they're the backdrop for the most useful strategy conversation I've read this month...
Anthropic disclosed the most comprehensive model distillation attack forensics to date. MiniMax generated 13M+ exchanges targeting agentic coding and tool use. Moonshot AI targeted agentic reasoning and computer-use agents (3.4M exchanges). DeepSeek focused on reasoning and ce...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.