HAWK Withdrawn From NIST After Claude’s Cryptanalysis
The fallout from this week’s cryptanalysis is now official: the HAWK team has withdrawn the scheme from NIST’s additional post-quantum signature standardization process. After reviewing the attack Claude Mythos Preview derived against the Lattice Isomorphism Problem at HAWK’s core — which roughly halves the block size needed in lattice reduction to recover an equivalent key — the authors confirmed the finding and pulled the candidate. It will not be standardized or deployed.
Two caveats matter. This is not a polynomial-time break, and Anthropic says it does not extend to other NIST signature candidates or to lattice cryptography broadly. Crucially, the already-finalized standards — ML-KEM and ML-DSA — rest on different math and remain secure and ready to deploy. Still, it’s the first time an AI-discovered weakness has removed a scheme from the running.
Claude Enterprise Ships an Admin API and Spend Controls
Anthropic is tackling the “surprise AI bill” problem head-on. Claude Enterprise now includes model-level entitlements, spend alerts, and an Admin API (in beta), giving IT and finance teams a way to cap and monitor token spend before it runs away as agent usage scales.
The refreshed admin analytics dashboard breaks down usage and cost by group and by user, showing concrete output — artifacts created, files edited, skills and connectors used — right next to what each one costs. The Admin API also handles member management: listing people, changing roles, removing members, and managing groups programmatically.
Cowork Comes to Mobile and Web With Offline Background Tasks
Claude Cowork — the agentic workspace for non-developers — now runs on iOS, Android, and the browser, with sessions and files following you across devices. On web it launches straight from the claude.ai home screen; on mobile it lives in the sidebar of the Claude app, rolling out to Max subscribers in beta.
The headline feature is cloud background execution: scheduled tasks keep running with no device online. Set a 6 AM client-prep task and Claude will work through email threads, transcripts, and recent news, build a briefing, and leave a follow-up drafted but unsent — then push a notification when it needs a decision or approval. Scheduled runs get the same connected tools, skills, and plugins as live sessions.
Subagents Go Three Deep, Run in the Background, and Feed Live Artifacts
Recent Claude Code releases keep pushing on agent orchestration. Subagents can now spawn nested subagents up to depth 3 by default (previously 1), and they run in the background so the main session keeps working and gets notified when they finish. Opus 5 is the default model, bringing a 1M-token context window and up to 128K output tokens.
Published artifacts can now pull live MCP connector data on every view, so a shared dashboard shows current numbers and can take actions on demand instead of freezing a snapshot. The claude-api skill now defaults to Opus 5 with a clean migration path from Opus 4.8.
Claude Science Opens Up to Every Paid Subscriber
Claude Science — Anthropic’s research product that aims to do for the lab bench what Claude Code did for the codebase — is now available to all paid Claude subscribers. It bundles more than 60 scientific databases and computation tools into one workspace and can autonomously carry out meaningful computational-biology and drug-development work from high-level instructions.
Alongside the product, Anthropic is running its own internal drug-discovery program, deliberately targeting “neglected” diseases that traditional biopharma tends to skip. It’s a notable widening of Claude’s footprint beyond software into frontier scientific work.
Claude Runs on NVIDIA GB300 Inside Microsoft Foundry
The enterprise cloud story keeps filling in. With Claude now generally available in Microsoft Foundry, NVIDIA highlighted that select models run on its GB300 GPUs behind Microsoft’s environment — customers deploy through Azure while Anthropic handles inference, using existing Azure auth, billing, and governance.
It reinforces the week’s throughline: meet enterprises where they already run. Between Foundry, the Cognizant alliance, stateless MCP, and now spend controls, Anthropic keeps stripping out reasons for risk-averse buyers not to standardize on Claude.
An AI Just Changed the Shape of a Standards Process
Yesterday the story was that Claude found a weakness in HAWK. Today it’s that HAWK is gone — withdrawn from NIST’s process by its own authors. That’s a different kind of milestone. A model didn’t just publish an interesting result; it altered which cryptographic scheme the world will be allowed to standardize on. The feedback loop from “AI cryptanalysis” to “real standards decision” just closed in a matter of days.
The reassuring footnote — ML-KEM and ML-DSA are untouched, and this isn’t a general lattice break — is real and worth repeating. But the process lesson is bigger than any one algorithm: standards bodies now have a tool that can pressure-test every candidate cheaply and repeatedly, and the prudent move is to run it against everything before ratification, not after. Better to lose a scheme in review than to standardize one an attacker can later unravel.