Thursday, July 30, 2026

The governance side moved too: 1,178 frontier-lab employees asked Washington to build a "pacing" mechanism, and OpenAI and Anthropic endorsed it as companies. And a maximum-severity flaw in the Ruflo (ex-Claude-Flow) agent harness landed a CVSSCommon Vulnerability Scoring SystemA standard 0 to 10 severity score: 0.1–3.9 low, 4.0–6.9 medium, 7.0–8.9 high, 9.0–10 critical.FIRST ↗ 10.0.

TLDR

Anthropic's Claude Mythos finds new cryptographic weaknesses — halves the effective key strength of the post-quantum scheme HAWK and speeds a 7-round AES attack 200–800x; disclosed to authors/NIST, no production impact (Jul 28) Security Trending
1,178 AI-lab staff call for a government "pacing" mechanism — OpenAI and Anthropic formally endorse the "Pacing the Frontier" letter asking the US to build tools to slow AI if needed later (Jul 28–29) Policy Trending
RufRoot: CVSS 10.0 RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ in the Ruflo agent harness — an unauthenticated MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ bridge exposing 233 tools (incl. shell exec) allowed RCE + AI-memory poisoning; patched in 3.16.3 (Jul 28–29) Security Agents
OpenAI opens ChatGPT for Academic Researchers — free GPT-5.6 Sol Pro access, scaling from 10K to 100K scientists by 2027, part of a $250M+ program (Jul 29) Customer AI
Meta + BlackRock form a $14B, 1GW data-center JV in El Paso — BlackRock funds own 80%; Meta contributes land/CIP and is sole initial tenant, online 2028 (Jul 28) Infra

Action Items

Things your engineering team should look at or do this week.

If you run Ruflo / Claude-Flow, upgrade to 3.16.3+ now and lock down the MCP bridge. Pre-3.16.3 Docker deployments exposed POST /mcp unauthenticated with terminal_execute reachable from the network — treat any exposed instance as compromised (rotate provider keys, inspect AgentDB memory for poisoned patterns). (The Hacker News | NVD)
Read Anthropic's cryptanalysis write-up as a capability signal, not a break. Nothing in production is affected and full AES holds, but "AI-assisted cryptanalysis as a compute line item" is now real — worth flagging if your roadmap bets on specific PQC candidates like HAWK. (Anthropic)
Check eligibility for OpenAI's academic program if anyone on your team collaborates with a university lab — it grants free GPT-5.6 Sol Pro and doesn't train on submitted data by default. (OpenAI)

Models & Tools

OpenAI launches ChatGPT for Academic Researchers (Jul 29) Customer AI — A program giving select academic researchers free access to GPT-5.6 Sol Pro plus hands-on support, starting at 10,000 participants this summer and scaling toward 100,000 by 2027 (each can invite four collaborators). Data isn't used for training by default; it's part of a $250M+ scientific-research push, with early users spanning genetic-disease diagnosis and astrophysics simulation. (OpenAI | Engadget)

Security

RufRoot — CVSS 10.0 unauthenticated RCE in the Ruflo agent harness (Jul 28–29) Security Agents — Noma Labs disclosed CVE-2026-59726 in Ruflo (formerly Claude Flow), a multi-agent orchestration harness for Claude Code and Codex. Its default Docker deployment exposed an MCP bridge (POST /mcp) with 233 tools — including terminal_execute, DB ops, and memory storage — to the network without auth, allowing an attacker to get a shell, steal provider API keys, and poison AgentDB "learning" memory (effects that can persist past mitigation). Fixed in 3.16.3 (locked-down default); maintainer patched within 24h of the Jun 30 disclosure, now public. (The Hacker News | Dark Reading)

Industry & Business

Meta and BlackRock form a $14B AI data-center JV in El Paso (Jul 28) Infra — A 1GW campus, online 2028, with BlackRock-managed funds owning ~80% and Meta ~~20%; Meta contributes land and construction-in-progress (~~$2.3B) and is the sole initial tenant, while BlackRock puts in ~$4.9B cash. Another datapoint in the shift to off-balance-sheet financing for hyperscale AI capacity. (Bloomberg | Forbes)
Freehand raises $75M Series B for supply-chain agents (Jul 29) Agents — The San Francisco startup ($100M total raised) deploys AI agents that negotiate with suppliers, process invoices, enforce contracts, and reconcile payments; customers include Meta, Unilever, J&J, Pfizer, and Dunkin'. Co-led by Battery Ventures and NewRoad Capital — a marker of enterprises paying for agents that act, not just answer. (Forbes | Crunchbase News)
Encore AI raises $30M Series A for revenue agents (Jul 29) Customer AI — The NY company (formerly Insait IO) mines calls, chats, emails, and CRM signals to learn what top sales/service reps do, then deploys agents to replicate it — pitched as revenue generation rather than headcount cuts. Led by Team8; several banks and insurers were customers before investing. (TechCrunch)

Policy & Safety

1,178 frontier-lab employees ask the US to build a "pacing" mechanism (Jul 28–29) Policy Trending — The "Pacing the Frontier" letter asks Washington to help develop the technical and governance tools to deliberately pace automated AI development — explicitly not a call to pause now, but to make a coordinated, verifiable slowdown possible later. Signatories include Dario Amodei, OpenAI's Jakub Pachocki and Mark Chen, Meta's Shengjia Zhao, and Google's Anca Dragan; OpenAI and Anthropic then endorsed it as companies. Follows the OpenAI-model-breaches-Hugging-Face incident. (Washington Post | CNN)

Paper Spotlight

Anthropic: discovering cryptographic weaknesses with Claude Mythos (Jul 28) Security — Using its unreleased Claude Mythos Preview, Anthropic improved the best-known attack on HAWK (a post-quantum signature scheme that survived two years of expert review) in ~60 hours — cutting the smallest config's key-recovery cost from 2⁶⁴ to 2³⁸ operations — and sped up a 7-round AES attack 200–800x, beating a 2013 record. Each result cost ~$100K in API usage and hundreds of hours of human verification; findings were disclosed to authors, NIST, and government/industry partners. No production systems are affected and full AES is untouched, but it's a concrete look at AI-assisted cryptanalysis. (Anthropic | The Hacker News)

Cryptanalysis became an AI capability story: Anthropic showed Claude Mythos Preview improving the best-known attacks on the post-quantum signature scheme HAWK and on round-reduced AES — years of expert review beaten in dozens of hours (no production system affected).