Claude Mythos Finds a Real Weakness in the HAWK Post-Quantum Cipher
Anthropic disclosed that Claude Mythos Preview uncovered a previously unknown structural weakness in HAWK, a lattice-based signature scheme and third-round candidate in NIST’s post-quantum standardization. By spotting an unexploited mathematical symmetry, it improved the best-known key-recovery attack on the smallest HAWK-256 parameter set from roughly 2⁶⁴ to 2³⁸ operations — in about 60 hours, something two years of expert human cryptanalysis had not managed.
In a separate run, Claude produced an improved meet-in-the-middle attack on 7-round reduced AES-128 that runs 200 to 800 times faster than the prior best, over three days of largely autonomous work and about a billion output tokens. Anthropic stresses neither result touches production systems — real-world AES uses the full 10 rounds — but it’s a striking demo of AI doing frontier math research.
The Stateless MCP 2026-07-28 Spec Comes to Claude
Anthropic is bringing the MCP 2026-07-28 specification to Claude, and it’s one of the biggest protocol releases yet. The core change: MCP moves from a bidirectional stateful protocol to a stateless request/response model, so servers can finally deploy on serverless and edge infrastructure instead of holding a persistent connection.
Authorization now aligns with production OAuth 2.0 and OIDC, so servers plug into enterprise identity systems like Entra or Okta without workarounds. Apps and Tasks ship as versioned extensions — a formal path to interactive UIs and long-running jobs via tasks/get, tasks/update, and tasks/cancel. MCP just passed 400M monthly SDK downloads, and Claude’s directory now lists 950+ connectors.
Published Artifacts Get Live MCP Data — and Subagents Go Three Deep
Recent Claude Code releases let a published artifact call MCP connectors every time someone views it, so a shared dashboard shows live data and can take actions on demand rather than a frozen snapshot from the session that built it. New sharing and collaboration options for Team and Enterprise, plus a screen-reader mode for accessible sessions, round out the update.
Under the hood, subagents can now spawn nested subagents up to depth 3 by default (with an env var to disable nesting), and Opus 5 is the default model with faster performance and expanded context. There’s also a background /code-review, safer auto mode and trust handling, and better MCP and Windows path handling.
What Stateless MCP Means for the People Actually Building Agents
The abstract spec change lands as a very concrete win for developers. A stateless core means you can host an MCP server as a plain Lambda or edge function — no long-lived socket, no sticky sessions, far cheaper to scale — and the new Tasks extension gives long-running work a first-class handle instead of a hacky polling loop.
Pair that with the platform’s recent mid-conversation tool changes and default fallbacks, and the direction is clear: Anthropic wants agent builders to compose capabilities dynamically, run them on commodity infra, and keep costs predictable. The connector directory crossing 950 servers shows the ecosystem is following.
Cognizant Deepens Its Anthropic Alliance, Trains 30,000 on Claude
Cognizant expanded its partnership with Anthropic, becoming one of a small number of Global Premier Partners in the Claude Partner Network. The move builds on a late-2025 tie-up and targets the gap between raw model capability and companies’ ability to turn it into business results, embedding Claude across the systems enterprises already run.
More than 30,000 Cognizant associates have completed Claude training under a new “Frontier Certified” workforce model. Cited wins include an agentic contract-intelligence system that cut biopharma review time up to 40% and an underwriter risk tool saving roughly eight hours a week — the kind of measurable outcome enterprises need to justify scaling AI.
Claude Settles Into Microsoft Foundry as the Azure Path Matures
Fresh on the heels of going generally available in Microsoft Foundry, Claude is becoming a standard Azure option: customers deploy select models through Microsoft’s environment while Anthropic runs inference behind them, using existing Azure authentication, billing, and governance. Inference runs in a U.S. data region to help with residency and compliance.
It fits the week’s theme — meeting enterprises where they already work. Between Foundry, the Cognizant alliance, and a stateless MCP that plugs into corporate identity, Anthropic keeps lowering the friction for risk-averse buyers to standardize on Claude without rebuilding their stack.
Cryptanalysis Just Became a Compute Line Item
The HAWK result is more than a headline. Two years of expert human review couldn’t find the symmetry Claude Mythos surfaced in 60 hours, and the AES speedup came from roughly a billion output tokens of largely autonomous work. Read that as a pricing statement: hard cryptanalysis — historically a scarce, artisanal skill — is starting to look like something you can buy in compute, scale up, and rerun.
That cuts both ways. Defenders get a tireless auditor that can pressure-test schemes before they ship; attackers, eventually, get the same tool. Neither result breaks a deployed system today, and that caveat matters. But if AI can meaningfully accelerate attacks on reduced-round primitives now, the prudent assumption is that the margin of safety on full ones just got a little thinner — and that post-quantum standards bodies should be running these models against every candidate, early and often.