AI Developer Digest

Wed, Jul 29, 2026

4 signals that cleared the gate38 scanned20 min read

The Signal — start here

Two threads run through today's digest. The first is the most significant: Claude Mythos Preview found a meaningful weakness in HAWK — a NIST post-quantum cryptography candidate — in roughly 60 hours, after two years of expert human cryptanalytic review found nothing. This is the first time an AI system has contributed a novel, confirmed result to post-quantum cryptography, and the research used a Claude Code scaffold as the experiment harness. The second thread is infrastructure aftermath from yesterday's MCP 2026-07-28 spec finalization: the Anthropic Python SDK shipped two hotfixes within one hour on July 28 to handle MCP SDK v2 compatibility (the v2 betas launched alongside the spec), and LiteLLM v1.94.0 rolled out with MCP OAuth improvements aligned to the new stateless spec. The spec finalization is immediately creating real developer friction — the migration window is not hypothetical.

Must-reads today

Anthropic Mythos finds HAWK post-quantum weakness — first AI discovery of a meaningful flaw in a NIST PQC candidate; found in ~60 hours what 2 years of human review missed; no production impact but a significant research signal for anyone in security or PQC

Anthropic SDK Python v0.120.2 — MCP SDK v2 compatibility hotfix on July 28; if you installed v0.120.1 and use anthropic[mcp] with MCP SDK v2, upgrade now

Breaking Changes

No breaking changes requiring forward-looking code changes this period. The Anthropic Python SDK v0.120.1 briefly restricted the mcp extra to <2 (breaking users who had installed MCP SDK v2), but v0.120.2 fixed this within one hour on July 28. If you are on v0.120.1, upgrade immediately — see API & SDK Changes for the full entry.

Model Releases

No model releases in the 24-hour window. Claude Opus 5 (July 24), Kimi K3 open weights (July 27, covered July 27), and vLLM v0.26.0 (July 27, covered July 27) are the most recent frontier events. No new open-weight or API models released July 28–29.

API & SDK Changes

Notable

Anthropic SDK Python v0.120.2 — MCP SDK v2/v1 Compatibility Fix

What changed

v0.120.1 (released 16:37 UTC) pinned the anthropic[mcp] extra to mcp<2, breaking any environment where MCP SDK v2 was already installed alongside the Anthropic SDK. v0.120.2 (released 17:38 UTC, one hour later) removed the strict pin and restored support for both MCP SDK v1 and v2 simultaneously.

TL;DR

Anthropic Python SDK v0.120.1 briefly broke anthropic[mcp] installs on MCP SDK v2 — fixed in v0.120.2 within one hour on July 28; upgrade immediately if you are pinned to v0.120.1.

Developer signal

Run pip show anthropic and verify you are on ≥0.120.2. If you pinned anthropic==0.120.1 in a lockfile (e.g., via pip-compile or poetry.lock), update and re-lock before your next deploy. The root cause: the MCP 2026-07-28 spec finalization shipped with MCP SDK v2 betas (Python, TypeScript, Go, C#), and the Anthropic SDK's mcp optional dependency hadn't been updated in time. v0.120.2 now supports both v1 and v2, so there is no forced migration to v2 today — v1 continues to work. If you are actively adopting MCP SDK v2 to experiment with the new stateless spec, v0.120.2 is the minimum Anthropic Python SDK version required. Note: MCP SDK v2 betas are still the migration path, not yet GA (~October 2026 expected); don't migrate production MCP servers to the v2 spec until the GA release.

Affects you ifYou use pip install 'anthropic[mcp]' and have (or plan to install) MCP SDK v2; you updated the Anthropic Python SDK on July 28 and landed on v0.120.1.EffortQuick — pip install 'anthropic[mcp]>=0.120.2' or update your lockfile.

anthropics/anthropic-sdk-python GitHub | Date: July 28, 2026 | Link: https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.120.2https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.120.2 | v0.120.1 context: https://github.com/anthropics/anthropic-sdk-python/releases/tag/v0.120.1

Research

High

Claude Mythos Preview Finds Real Weakness in HAWK Post-Quantum Signature Scheme

What changed

Claude Mythos Preview, operating within a Claude Code scaffold that allowed it to run automated cryptanalytic experiments, identified two new results: (1) a previously unknown structural weakness in HAWK — a lattice-based digital signature scheme and NIST PQC third-round candidate — by finding a nontrivial automorphism that enables a faster key enumeration attack, reducing effective security from approximately 2^64 to 2^38 operations; (2) an improved meet-in-the-middle attack on 7-round reduced AES-128 that runs 200–800× faster than the prior best-known attack.

TL;DR

Claude Mythos Preview found a HAWK (NIST PQC candidate) weakness that reduces security from ~2^64 to ~2^38 ops in roughly 60 hours — after two years of expert human review found nothing; also improved the AES round-reduced attack by 200–800×; no production system needs to change today.

Developer signal

No immediate code changes are required — HAWK is a candidate scheme, not deployed in production, and the AES finding targets a 7-round reduced variant (full AES-128 has 10 rounds and remains secure). Three concrete implications: (1) If you use HAWK in any cryptographic prototyping or protocol draft, monitor NIST's response; the 2^38 effective security level is within theoretical reach of large compute clusters and weakens HAWK's position as a PQC signature candidate. NIST's response timeline (weeks to months) will clarify whether HAWK's standardization path is affected. (2) If you are building AI-assisted security research tools, this is the first published evidence that a frontier model using a Claude Code scaffold can contribute a genuine novel result to post-quantum cryptography — the scaffold design (automated experiment runner, iterative hypothesis testing, result logging) is described in the research and is replicable. (3) The AES result has no practical implication for deployed systems — it requires 2^105 chosen plaintexts and would cost hundreds of millions of dollars to implement. The HAWK result is the operationally significant finding.

Affects you ifYou are evaluating or specifying post-quantum signature schemes that include HAWK; you are building AI-assisted security or cryptography research tools; you are following the NIST PQC standardization process.EffortQuick — no code changes required today; monitor NIST's response for HAWK-specific decisions.

Anthropic Research | Date: July 28, 2026 | Link: https://www.anthropic.com/research/discovering-cryptographic-weaknesseshttps://www.anthropic.com/research/discovering-cryptographic-weaknesses (returned 403 on direct fetch; details sourced from multiple corroborating secondary sources) | Corroborated by: https://www.techtimes.com/articles/321876/20260728/ai-cracks-post-quantum-cipher-60-hours-after-two-years-human-review-failed.htm, CyberScoop, Slashdot, explainx.ai summary

Tooling

Medium

LiteLLM v1.94.0 — Router Overhaul, MCP OAuth SSO, and Security Hardening

What changed

v1.94.0 (100+ merged PRs) adds adaptive complexity tiers to the LiteLLM router, session affinity options, and a plugin-based routing pipeline architecture. MCP OAuth gains interactive SSO sign-in and full token lifecycle management. Two security fixes address user budget enforcement on team keys (prior behavior did not correctly enforce per-user limits on team-scoped keys) and key logging sanitization. Docker images are now signed with cosign for supply chain verification.

TL;DR

LiteLLM v1.94.0 adds adaptive routing complexity tiers, MCP OAuth SSO, fixes user budget enforcement on team keys (security), adds GPT-5.6 and Gemini Omni Flash pricing support, and ships with Python 3.14 support and Docker cosign image signing.

Developer signal

Three items warrant verification after upgrading: (1) Budget enforcement fix — if you use team keys in LiteLLM and rely on per-user budget limits, the prior behavior was not correctly enforcing these limits on team-scoped keys. v1.94.0 fixes this; test your budget configuration after upgrading to verify limits apply as expected and that no previously uncapped users now hit unexpected limits in production. (2) MCP OAuth change — if you use LiteLLM's MCP proxy with OAuth in a non-interactive deployment (e.g., server-side only, no user browser interaction), the new interactive SSO sign-in flow may break your setup; audit your OAuth integration and confirm it works with the new token lifecycle management. (3) Docker cosign — verify the new image signatures if you have supply chain security requirements: cosign verify ghcr.io/berriai/litellm:v1.94.0. The adaptive routing tiers and plugin pipeline architecture are additive; no migration required, but the complexity tiers are now the recommended path for multi-model routing strategies that vary by task complexity.

Affects you ifYou use LiteLLM's team key budget enforcement; you use MCP OAuth with non-interactive (server-side only) deployments; you self-host LiteLLM via Docker with supply chain security requirements.EffortModerate — budget enforcement behavior changed (verify configs after upgrade), MCP OAuth flow updated (test non-interactive deployments), Docker cosign verification is new.

BerriAI/litellm GitHub | Date: July 28, 2026 | Link: https://github.com/BerriAI/litellm/releases/tag/v1.94.0https://github.com/BerriAI/litellm/releases/tag/v1.94.0

Benchmarks & Leaderboards

No new leaderboard movements in the 24-hour window. The SWE-bench Verified snapshot (Claude Opus 5 at 96%, Mythos 5 at 95.5%, Fable 5 at 95%) from the July 28 digest remains current. LMArena, Open LLM Leaderboard, and LiveCodeBench primary pages returned 403; no new July 29 snapshots available with a verifiable primary source.

Trends & Emerging Tech

AI-Assisted Cryptanalysis Is Becoming Real Scientific Work

What's happening

Claude Mythos Preview didn't assist a human cryptanalyst — it ran an autonomous hypothesis loop and found a novel attack that two years of expert human review hadn't discovered. The scaffold design (Claude Code as experiment harness, automated experiment runner, iterative hypothesis testing with result logging) is the key technical detail: it gave the model an automated feedback signal to iterate on cryptanalytic approaches without human supervision at each step. The HAWK finding (2^64 → 2^38 effective security) is a genuine advance, not a restatement of known results. Separately: the July 20 HuggingFace incident (near-miss, outside window) showed the inverse — an autonomous agent framework used without safety restrictions to breach production infrastructure. Both events in the same week are shaping the emerging debate about what frontier agentic models can and can't be constrained to do.

Why watch this

The research scaffold design — automated experiment runner, tight feedback loop, formal domain with clear success criteria — is domain-agnostic. Formal domains with automated execution environments (symbolic math, formal verification, exploit development, materials binding affinity) now have an existence proof that this approach works at the frontier level. If you are building AI-assisted scientific research tools, the Anthropic paper is the first peer-reviewed template. The practical near-term implication: expect NIST to respond to the HAWK finding, and expect more similar papers from other labs over the next 6–12 months as Claude Code–style scaffolds are applied to other formal domains.

Anthropic Research | Date: July 28, 2026 | Link: https://www.anthropic.com/research/discovering-cryptographic-weaknesses

Technical Discussions

Nothing cleared the quality bar this period. The HuggingFace July 20 security incident generated a high-signal HN thread (score >500, detailed technical discussion of the attack chain) but the primary disclosure is outside the strict 24-hour window — see Near-misses below for details and the developer action required.

Quick Hits

llama.cpp b10174 (July 29, 2026) — MTP (Multi-Token Prediction) speculative decoding support for GLM-5.2 models; enables draft-model-based fast generation for this architecture on all backends. https://github.com/ggml-org/llama.cpp/releases/tag/b10174
llama.cpp b10181 (July 29, 2026) — Disabled MMQ (quantized matrix multiplication) on NVIDIA CUDA devices reporting <48 KiB shared memory per block; falls back to BLAS path and prevents crashes on Moore Threads MTT S70 and similar hardware. https://github.com/ggml-org/llama.cpp/releases/tag/b10181
llama.cpp b10180 (July 29, 2026) — Adds contiguous fast path and 32-bit index math for unary element-wise SYCL operations; performance optimization for Intel GPU (SYCL backend) inference. https://github.com/ggml-org/llama.cpp/releases/tag/b10180
llama.cpp b10182 (July 29, 2026) — Reorganizes token suppression handling into the common sampling module and addresses security issues in the same change. https://github.com/ggml-org/llama.cpp/releases/tag/b10182
llama.cpp b10175 (July 29, 2026) — Separates RDNA 3 and RDNA 3.5 GPU architecture configurations for MMQ tuning parameters; enables independent performance tuning for the two AMD RDNA variants. https://github.com/ggml-org/llama.cpp/releases/tag/b10175
GitHub Copilot enterprise metrics impact dashboard (July 28, 2026) — New dashboard for enterprise admins and org owners showing adoption depth, usage cohorts, and developer productivity trends. Note: github.blog returned 403 on direct fetch; included from search snippet. https://github.blog/changelog/

Worth Watching (Announced, Not Yet Shipped)

Item	ETA	Notes
Anthropic Workbench sunset + experimental prompt tools retirement	August 17, 2026	`/v1/experimental/generate_prompt`, `/v1/experimental/improve_prompt`, `/v1/experimental/templatize_prompt` will error after August 17. Export saved prompts from the banner or Organizational Settings.
Claude Opus 4.1 retirement	August 5, 2026	Migrate to Opus 4.8 or Opus 5.
OpenAI o3 retirement	August 26, 2026	Migrate to GPT-5.6 Sol/Terra/Luna.
Imagen 4 shutdown	August 17, 2026	Migrate to Imagen 4 Ultra or Veo 3.
MCP Tier 1 SDK GA (estimated)	~October 2026	Python, TypeScript, Go, C# SDKs are in beta for 2026-07-28 spec; GA expected after ~10-week validation window. Rust SDK in beta support with no GA ETA.

Filtered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.