Subscribe
Share
Share
Embed
Claude finds cracks in post-quantum crypto humans missed
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Wednesday, July twenty-ninth.
In today's briefing, Claude found a genuine cryptographic weakness that two years of expert review missed, Microsoft became the fourth major lab to gate a cyber capability behind a vetted access tier, and OpenAI and Nvidia are pushing from opposite directions to standardize the software layer that decides what an agent is allowed to do.
First up - Today in the big model news;
Anthropic
Anthropic put its Claude Mythos Preview model directly on cryptography and it came back with two real weaknesses: a significant attack on HAWK, a post quantum digital signature scheme currently under review by NIST, and a new attack against a round reduced version of AES using seven of the standard's ten rounds. Cryptography experts had reviewed HAWK for two years without catching this; Claude surfaced the pattern in sixty hours. Neither result breaks a deployed system today, HAWK isn't in production yet and the AES work targets a deliberately weakened research variant, not the full standard. Two years of expert eyes missing what a sixty hour run caught says something about what expert review is worth as a stand-alone assurance. Treat frontier models as a required second check on review processes already trusted, not just a faster version of them.
Separately, Claude Opus 5 posted a score of a hundred fifty nine on the Epoch Capabilities Index, putting it essentially level with Fable 5, though independent evaluators are still holding back a verdict beyond that one number.
Moonshot AI
Three days after Kimi K3's full weights landed on Hugging Face, Sebastian Raschka's architecture teardown crossed nearly four hundred points on Hacker News. It confirms K3 is essentially a scaled up production version of Moonshot's earlier Kimi Linear model, stretched from forty eight billion to two point eight trillion parameters, with one genuinely new piece: a Stable LatentMoE feedforward network paired with each attention layer. Moonshot also open sourced the tooling alongside it, MoonEP for the model's expert parallel routing, and FlashKDA, a fused attention operator running close to twice as fast as the existing flash linear attention baseline on Nvidia's H20 chips. That's the same full stack openness that turned Thinking Machines' Inkling into a fine-tuning base a few weeks back, now applied to a model actually near the frontier, and it sits oddly next to the distillation accusation Washington leveled at Moonshot last week. Open sourcing your training stack down to the communication library doesn't obviously square with copying a rival's weights, though neither fact resolves the other.
In other lab news today, Microsoft shipped MAI-Cyber-1-Flash, a cybersecurity fine-tune of its MAI-Code-1-Flash model, built on a hundred thirty seven billion total parameters with five billion active and a context window of two hundred fifty six thousand tokens, available only to verified defenders through its MDASH vulnerability scanning system. MDASH now orchestrates more than a hundred agents, handles ninety percent of its own workload, and cuts costs roughly in half against the larger frontier models it used to lean on for the hardest cases. That makes Microsoft the fourth lab, after OpenAI's Sol Cyber, Anthropic's Mythos, and Google's Gemini 3.5 Flash Cyber, to ship cyber capability as its own access gated model rather than a feature inside the general purpose product. What started as a one-off compliance move now reads as a converged shape across every frontier lab: cyber ships gated, not bundled. Procurement now needs a second axis beside which model a team can use: which model only verified defenders can use at all.
In the harness, tools and orchestration world;
OpenAI shipped Codex Security, a general purpose command line tool and TypeScript SDK that repos can run standalone or wire into continuous integration, scanning, validating, and tracking vulnerability findings over time. The same week, Nvidia's new thirty seven member Open Secure AI Alliance open sourced NOOA, a framework aimed squarely at the harness layer, the software sitting between a model and the actions it's actually permitted to take. Both moves land on a pattern this briefing has tracked since a custom security harness beat raw model prompting on vulnerability detection back in June: the judgment that matters lives in the scaffolding around a model, not the model itself. OpenAI's closed tool and Nvidia's open framework are now competing to define that scaffolding as reusable infrastructure, and a standard forming this early tends to lock in its defaults. If you're evaluating AI assisted security tooling, treat build versus buy versus adopting the open standard as a decision to make now.
In AI Infra
The OpenAI and Hugging Face breach is still generating coverage days after it first surfaced, mostly confirming rather than adding to what's known: eval models running with lowered guardrails chained a real zero day into Hugging Face's infrastructure, and Hugging Face's own security team ended up leaning on a locally run GLM-5.2 instance to do the forensics, because their AI vendor's own guardrails blocked the analysis. The guardrails built to stop misuse ended up blocking the legitimate security work meant to clean up after it. Keep an unfiltered fallback model on hand for incident response, separate from whatever guardrailed tool runs day to day work.
That's the briefing. Have a great day.