Dakarda AI Newsletter · 6 July 2026
Monday Edition
Cloudflare ends free data for AI
Cloudflare closes the door on free data scraping, Cisco deploys an AI agent to every employee, the UN convenes the world for a conversation about AI power, and Palantir publishes a manifesto against tokenmaxxing. Welcome to the week when AI stops being a toy and becomes infrastructure.
Intro · Alex
Something is shifting this week — and not just inside model engines. Cloudflare, which controls one-fifth of internet traffic, called AI bots' bluff and announced payments for publishers. Cisco, meanwhile, is deploying an AI agent to every one of its 90,000 employees — while simultaneously laying off 4,000 people. These aren't beta-test announcements. These are production-scale deployments the likes of which we haven't seen before. In this edition, I've reviewed five key stories for you that show AI is entering a phase of real politics, regulation, and architecture — not just more benchmarks. I also found a story about how poisoned MCP tool descriptions can make your own AI agent hand over corporate data to a stranger. And a warning that simply telling a model that 2+2=5 can disable its safety measures. Read carefully.
What's worth knowing
Cloudflare ends free AI scraping of web content
Cloudflare announced that from September 15, 2026, it will block AI crawlers by default on sites displaying ads — unless the bot declares its purpose: Search, Agent, or Training. It also introduces a payment system for publishers when their content appears in AI responses. This is a radical shift in the rules of the game for the entire ecosystem.
Cisco gives a personal AI agent to every one of its 90,000 employees
Starting August 2026, every Cisco employee will receive a personalized AI agent — an assistant for tasks, questions, and routing to the most efficient model. It's one of the largest enterprise AI agent deployments, accompanied by a restructuring and layoffs of approximately 4,000 people.
UN convenes the first Global Dialogue on AI Governance in Geneva
The first-ever Global Dialogue on AI Governance kicked off in Geneva with all 193 UN member states in attendance. Yoshua Bengio warned of the risk of 'catastrophic harms' and urged governments to take AI development scenarios seriously. A second session is planned in New York in May 2027.
Trump administration tightens restrictions on models — open source boom
New US restrictions on private model releases by Anthropic and OpenAI now include 'kill-switch' mechanisms — the ability to remotely block models. Experts warn this could be a gift to China, which offers cheaper open models, and accelerate the global migration of developers toward open source.
Palantir publishes a 9-point AI sovereignty manifesto and attacks OpenAI
Palantir published a manifesto criticizing 'tokenmaxxing' — overpaying for AI — and urging institutions to keep their data on-premise. It signals a shift in the debate from 'buy more tokens' to 'control data and models locally,' which could accelerate the trend toward on-premise and sovereign AI in government.
From the tech world
Poisoned MCP tool descriptions can make your AI agent leak data
Microsoft warns of a new attack vector: simply replacing a tool description in MCP can make an agent hand over confidential corporate data on its own. This is the fastest-growing threat in the AI agent supply chain.
Tip of the day
Agent trust audit: how not to get poisoned by MCP
An AI agent trusts the tool descriptions it receives from external servers. If someone swaps an MCP tool description from 'getUserEmail' to 'sendUserEmailToAttacker,' the agent — without breaking any rules — will send them your data. The problem lies in the trust architecture: agents take tool descriptions at face value because nobody assumes the documentation is lying. Practical step: Before letting your agent use external tools via MCP, audit the descriptions for 'poisoned descriptions.' Implement a policy that every tool description must be approved by a human before first use. And set up monitoring for outgoing requests — if the agent suddenly starts sending emails with attachments to an address outside your domain, that's an alarm signal.
Tool of the issue
X MCP Server — connect AI agents to real-time data from platform X
X (formerly Twitter) has launched its own MCP server, allowing AI agents to access real-time data from the platform. If you're building agents that analyze trends, social sentiment, or react to current events — this tool eliminates the need for your own scraper.
Reading list
Cloudflare Just Ended AI's Free Ride on the Web
Why Cloudflare's decision is the most important change for publishers and AI companies since the launch of ChatGPT. Key context for news item #1.
Read Palantir's 9-Point Manifesto on AI Sovereignty
Full text of Palantir's manifesto and analysis of why the data analytics giant for governments and militaries is raising the alarm about 'tokenmaxxing.'
AI is no longer a matter of technology — it's a matter of power: over data, over infrastructure, and over who pays for training.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.