AI Intelligence Report
Top Stories
Pentagon strikes classified-network AI deals with eight Big Tech firms — and freezes out Anthropic
The U.S. Department of Defense has finalised agreements with OpenAI, Google, Microsoft, Amazon Web Services, Oracle, Nvidia, SpaceX and Reflection AI to deploy their AI on classified secret and top-secret networks. Anthropic — a leading frontier AI lab — was excluded after the Pentagon designated it a “supply-chain risk” in March, a label normally reserved for foreign adversaries like Huawei. The dispute traces back to Anthropic’s refusal to drop two firm conditions on Defense use: no autonomous weapons targeting and no domestic mass surveillance. The two sides are now in litigation, and the designation also limits how defence contractors can use Claude in their stacks even on non-classified work.
Anthropic and OpenAI both launch joint ventures for enterprise AI services
In a notable parallel move, both Anthropic and OpenAI announced separate joint ventures focused on deploying enterprise AI services. Anthropic’s vehicle includes named investors TPG, Brookfield Asset Management, Advent and Bain Capital — a clear signal that the frontier labs are racing to capture corporate budgets directly rather than only through cloud partners. The structure also gives both companies new ways to bundle services, deployment and consulting around their models, and ratchets up competition with Microsoft, Google and the systems integrators.
GPT-5.5 enterprise rollout continues; broader access expected by August 14
OpenAI’s GPT-5.5 — codenamed “Spud” and pitched as the company’s most capable model for professional work — is rolling out further to Plus, Pro, Business and Enterprise users in ChatGPT and Codex, with Codex offering a 400K context window. The model claims meaningful gains in agentic coding, computer use and multi-step research, and matches the previous version’s per-token latency. A Fast mode generates tokens 1.5x faster at 2.5x the cost. Pricing has settled at roughly $2.25 per million input tokens — making frontier-grade AI cheaper than enterprise software lines that cost thousands of dollars a month two years ago.
Anthropic’s Mythos cybersecurity tool sparks Washington alarm
Anthropic’s Claude Mythos Preview — distributed only through the restricted Project Glasswing programme to Amazon, Apple, Google, Microsoft, Nvidia, JPMorgan Chase and a handful of others — has been demonstrating an 83% first-attempt success rate at exploit creation, and has already discovered thousands of high-severity vulnerabilities in every major operating system and web browser. CEO Dario Amodei visited the White House following the unveiling, and former U.S. National Cyber Director Kemba Walden told Fortune that critical-infrastructure operators “aren’t ready” for what AI-assisted hacking can now do. European finance ministers are publicly demanding access. One reason this matters for comms: it is the clearest example yet of an AI capability advancing faster than any country’s disclosure or coordination playbook.
Tencent doubles down on agentic AI with latest Hunyuan updates
Coverage continues of Tencent’s pivot to an “agent-first” AI strategy. The Tencent Cloud Agent Development Platform (TCADP) — rebranded from the former knowledge engine — is being positioned as a one-stop platform for building enterprise-grade AI agents, integrating Tencent Cloud’s RAG framework with tools to activate domain-specific knowledge. This will be relevant context for the broader Hunyuan ecosystem upgrade due later this month, when fast-thinking Hunyuan Turbo S, deep-thinking Hunyuan T1, T1-Vision, Hunyuan Voice, Hunyuan Image 2.0, Hunyuan 3D 2.5 and Hunyuan-Game are expected to be unveiled together.
AI News Roundup
Substack Highlights
Inoreader AI Folder
Microsoft VS Code says Copilot AI wrote all your code
In Visual Studio Code versions 1.117 and 1.118, any use of autocomplete (including ordinary tab-completion) gets stamped with a “Co-authored by Copilot” commit trailer — even with the AI features setting explicitly disabled. Developers are reporting they checked their commit messages before committing and the trailer was added anyway. A small bug, but a striking attribution problem.
How OpenAI delivers low-latency voice AI at scale
Engineering deep-dive on how OpenAI rebuilt its WebRTC stack for real-time voice with low latency, global scale, and seamless conversational turn-taking. Important context for anyone planning voice agents in customer service, podcast production or media interviews.
Meta is moving into humanoid robots
Meta is committing significant capital and engineering talent to a humanoid robot programme, joining Tesla, Figure and a growing list of Chinese players. The newsletter pairs the news with the Pentagon’s Anthropic exclusion as a reminder that frontier AI now also lives in physical hardware.
Code with agents (without breaking things)
A practical guide to running coding agents on real codebases. The bigger lesson for non-engineers: keep agents on a tight scope, build in review checkpoints, and isolate runs so a bad agent action can’t take everything else with it. The same template applies to PR/comms work.
The Prompt That Automates Your Job
A how-to on building a reusable role-specific AI prompt library — the kind a senior comms director could maintain alongside their style guide. Includes templates for repetitive briefs, executive summaries and inbound-press triage.
Patch now or pay later
Frames an oncoming software patch wave that follows the Mythos-class disclosure regime, plus a colourful aside that GPT-5.5 has been used to plan its own launch party — agentic capability creeping out of demo land.
Harvard puts out a new study that has doctors sweating
New Harvard study finds AI-assisted diagnosis matching or beating doctors on a meaningfully wider class of conditions. The same edition asks whether layoffs blamed on AI are really driven by AI or by executives using AI as cover for cuts.
Data to start your week: AI boom, nowhere near the ceiling
Azeem Azhar’s data note: the “compute crunch” is real, but a meaningful share of supply is still latent — racks installed, capacity provisioned, waiting on enterprise budgets to unlock spend. The under-investment risk, he argues, still outweighs the over-investment one.
Anthropic is getting pulled directly into Wall Street
Hedge funds, trading desks and bulge-bracket banks are increasingly building Claude into core research workflows — a notable counterpoint to the Pentagon-Anthropic ban. The financial sector’s appetite for Claude appears to be growing despite the federal posture.
LWiAI Podcast #243 — GPT-5.5, DeepSeek V4, AI safety sabotage
Latest podcast covers GPT-5.5’s enterprise rollout, DeepSeek V4’s release in China, and a new line of research on AI safety sabotage — adversarial attempts to undermine alignment training during fine-tuning.
Starbucks just rolled back its automation push
Starbucks has reversed parts of its store-automation rollout, citing customer experience concerns. The same edition runs an interview with UChicago economist Alex Imas on what becomes scarce — judgment, taste, accountability — when AI makes everything else cheap.
Open Up: Anthropic launches Claude Security for Enterprise; Meta opens ad accounts to AI agents
Two structurally important moves: Anthropic launches Claude Security as an enterprise-only product, and Meta opens advertiser accounts to AI agents — a real shift in how digital ad budgets might be deployed. Also flags Perplexity’s new financial research tools.
Gemini Gets to Work, Claude’s Big Pull, and OpenAI Unchained
Sketches the current state of the AI race as a three-way fight: Gemini moving into actual enterprise production, Anthropic pulling enterprise contracts, and OpenAI loosening Microsoft constraints to broaden its commercial reach.
Exponential View #572: When certainty is the real AI risk
From a week of meetings in China with Zhipu, MiniMax, Kimi, Alibaba, Xiaomi and ByteDance: Azhar’s argument is that the most dangerous trait in current AI is unwarranted certainty — confidently wrong outputs at scale. A useful mental model for risk and crisis comms.
Gemini and new evidence on the Declaration of Independence copperplate
Kalev Leetaru uses Google’s Gemini to do high-resolution comparative analysis of NARA’s vellum and paper Stone copies of the Declaration of Independence — finding new evidence that both came from the same copperplate. A nice example of AI-assisted document forensics; the same workflow translates to comms-document provenance work.
Tokenmaxxing: how much did you spend in tokens?
David Gerard’s piece (here, in a Portuguese translation) skewers the rising “token spend” vanity metric. A useful corrective when AI-vendor sales pitches lean too hard on usage volume as proof of value.
AI Workflows & Tool Watch
Claude Code’s new alwaysLoad MCP option and auto-retry logic
Anthropic added an alwaysLoad setting to Claude Code’s MCP server config — when enabled, all tools from that server skip the tool-search deferral and are always available in the session. MCP servers also now auto-retry up to three times on transient errors (5xx, connection refused, timeouts) before being marked as failed. This means more reliable Cowork sessions when you’re stitching together connectors for, say, Slack + Outlook + WordPress in a single workflow.
Anthropic ships another wave of Cowork plugins for specific job functions
Anthropic released 12 more official Cowork plugins covering HR, Engineering, Design, Operations, Financial Services and more — on top of the original 11 from launch. Admins can also now create private plugin marketplaces as part of enterprise deployment. For a global comms function, this is the right moment to consider a private “Tencent Comms” marketplace bundling your skills and connectors so the team has a consistent setup.
Perplexity Personal Computer turns a Mac mini into an always-on AI agent
Perplexity has rolled out Personal Computer to Max subscribers and earlier waitlist members. It runs continuously on a dedicated Mac mini, integrating directly with local files, native apps and your existing workflows — voice orchestration included, and now using Claude Sonnet 4.6 / Gemini 3.1 Pro / Opus 4.7 under the hood. Worth piloting alongside your Eagle/DEVONthink/Obsidian stack as a passive monitor for media mentions and inbound press.
Read AI’s MCP server brings meeting transcripts straight into Claude
An XDA write-up details how connecting Read AI’s MCP server to Claude completely changes a meeting workflow: transcripts, action items and summaries become directly queryable in Claude itself, no more hunting in a separate tool. For anyone running a global comms team across time zones, this is a credible alternative to managing Zoom + VooV + Otter + Notion separately.
n8n’s AI Agent node + LangChain put a real agent at the centre of your workflow
n8n now ships a native AI Agent node: an LLM decides which connected tools to call, reads the results, and chains actions until the task is done. Combined with the matured LangChain integration, this turns n8n from “a workflow builder with AI steps” into “an agentic workflow builder.” For comms automation — newsroom monitoring, draft generation, escalation routing — this is the most flexible self-hosted option right now.
Obsidian Skills: official AI plug-in framework for vaults
Obsidian CEO Steph Ango open-sourced “Obsidian Skills” — a skill pack that teaches AI assistants how to use Obsidian properly. It includes things like a Canvas skill that drops generated canvases directly into a Drafts folder in your vault. The cleanest way yet to let Claude or Gemini operate over your existing PKM rather than a parallel one.
Practical pattern: a reusable role-specific prompt library
Bagel Bots’ “Prompt That Automates Your Job” argues for treating prompts as a maintained library — versioned alongside your style guide, with templates for repetitive briefs, executive summaries, inbound-press triage and crisis Q&A. Pairs cleanly with Drafts (for capture) and Obsidian (for storage and retrieval).
Tencent Mentions
Tencent doubles down on agentic AI with latest Hunyuan updates
Coverage of Tencent’s pivot to an agent-first AI strategy and the Tencent Cloud Agent Development Platform (TCADP), positioned as a one-stop platform for building enterprise-grade AI agents on top of the Hunyuan stack. Useful framing in advance of the broader Hunyuan ecosystem upgrade later this month.
Tencent unveils Hy3 Preview: agent-first model from the Hunyuan rebuild
Tencent’s official site continues to push Hy3 Preview as the first model from the 2026 Hunyuan rebuild, with messaging centred on agent capabilities and real-world usability. Worth tracking how this lands in Western trade press over the coming days.
Exponential View flags Tencent in its China AI sweep
Azeem Azhar’s most recent edition mentions a China tour covering Zhipu, MiniMax, Kimi, Alibaba, Xiaomi, ByteDance and others — Tencent did not appear in the named visit list, which is itself a small signal worth noting. The Western analyst circuit’s mental map of Chinese AI still under-weights Tencent relative to the Hunyuan platform’s actual scope.
WeChat AI agent in development to rival OpenAI
Earlier coverage flagged a planned WeChat-integrated AI agent designed to compete with OpenAI’s consumer surface. Watch this space — any consumer-grade agent baked into WeChat is a story Western tech press will want to cover.