|
|
|
Anthropic just made the AI infrastructure boom feel very, very real. | TeraWulf announced a 20-year lease with Anthropic at its Justified Data campus in Hawesville, Kentucky, expected to generate roughly $19B in contracted revenue. The campus is planned for about 401 MW of critical IT load, with first capacity coming online in late 2027 and full ramp by early 2028. | Translation: AI companies are no longer just buying chips. They are effectively reserving towns, power loads, and decade-long chunks of the electrical grid. At this point, “we need more compute” has gone from startup pitch deck line to “please build us a small industrial moon base in Kentucky.” | Here's what happened in AI today: | 😼 A patched Google Dialogflow flaw showed how one rogue AI agent could hijack enterprise chatbots. 📰 Illinois signed a frontier AI safety law. 📰 China reportedly considered restricting overseas access to top models. 🍪 Meta rolled out Muse Image across Instagram, WhatsApp, and Meta AI. 📖 Nature published The AI Scientist research.
| Hey: Want to reach 700,000+ AI-hungry readers? Advertise with us! | Love robots? We just launched a robotics newsletter! Sign up for it here. | | 😺 A Rogue Google Dialogflow Agent Could Hijack Enterprise Chatbots | You know how every company wants an AI chatbot that can talk to customers, answer questions, and connect to real business systems? | Great. Now imagine that chatbot quietly starts reading conversations, stealing credentials, and impersonating itself. | That is the lovely little nightmare Varonis Threat Labs disclosed this week in Google Cloud’s Dialogflow CX, Google’s platform for building customer-service chatbots and voice agents. | Here’s what happened: | Varonis found a vulnerability it named Rogue Agent inside Dialogflow CX’s Code Blocks feature. Code Blocks let developers add custom Python logic to chatbot workflows. With one edit permission on one agent, an attacker could inject malicious code into the agent pipeline. Google issued an initial fix in April and fully resolved the issue in June. Varonis said it is not aware of any real-world exploitation before the patch.
| The core issue was not “the AI got tricked.” It was worse and more boring: the plumbing around the AI trusted the wrong thing. | Dialogflow Code Blocks ran inside a Google-managed Cloud Run environment. According to Varonis, agents in the same GCP project effectively shared that environment, and attackers could override a key execution file. That gave them access to conversation history, session details, and the ability to force the chatbot to return attacker-chosen messages. | In plain English: a compromised bot could quietly ask users to “reauthenticate,” collect credentials, and make the whole thing look like a normal chatbot response. Customer support, but with Ocean’s Eleven energy. | Why this matters: Enterprise AI agents are quickly moving from “answer this FAQ” to “touch customer data, trigger workflows, and talk to backend systems.” That makes agent permissions the new security boundary. If one lightly protected workflow can see too much, write too much, or share too much runtime with other agents, the chatbot becomes a front door with a very confident welcome mat. | Varonis and Google recommend auditing Dialogflow CX configurations for suspicious Playbook updates, reviewing past update actions, and checking Code Blocks for unauthorized changes. | Our take: The second-order lesson is that agent security will not be solved by better prompts. It will be solved by boring controls: narrow permissions, isolated runtimes, visible logs, and default skepticism toward any agent that can execute code. | The next enterprise AI race is not just who can deploy agents fastest. It is who can prove those agents cannot quietly become the intern from hell with production access. | |
|
|
| Plenty of companies can launch an AI pilot. Far fewer know how to make it stick. Explore this resource hub, sponsored by Dell AI Factory with NVIDIA, for strategies, decisions, and real-world lessons on turning AI into something scalable, useful, and worth the investment. | Learn More | |
🎓 AI Skill of the Day: Make Fable Stop Reading Junk | Expensive AI sessions usually do not fail because the model is dumb. They fail because you made it read a junk drawer. | After spending $1,486 testing Fable tokens, Nick Saraev landed on one rule: token management is context management. Before a long Claude Code / Fable run, make the model read less irrelevant stuff. | Here’s the workflow: | Compress your system prompt and memory files so they preserve meaning with fewer words. Tell the model to search before reading giant files. Put logs, CSVs, or big datasets behind a query tool instead of pasting raw text. Default “thinking” to low, then raise it only for hard decisions. Use /context checks to catch hidden bloat from tools, skills, or MCPs.
| Nick’s secondary trick is turning giant static prompts into compact image references for repeated runs, but treat that as experimental and verify quality. The safer universal win: make the model look only where the answer probably lives. | Audit this AI workflow for context waste.
Task:
[describe the thing I want the model to do]
Current context:
[paste system prompt, project instructions, file list, logs, or workflow notes]
Please:
1. Identify what context is truly needed.
2. Flag anything bulky, repeated, irrelevant, or risky to read in full.
3. Rewrite my instructions with semantic compression: preserve meaning, remove filler.
4. Add context-frugality rules:
- Search before reading large files.
- Read specific regions, not entire files or folders.
- Use database/query tools for logs, CSVs, and tables.
- Ask before expanding beyond 3 files.
- Summarize findings before reading more.
5. Recommend the lowest thinking level that should work.
6. Give me a final copy-pasteable version of the optimized instructions.
| Want more tips like this? Check out our AI Skill of the Day Digest for July. | Have a specific skill you want to learn? Request it here. | | |
|
| The Neuron Academy is the self-paced learning library we built for readers like you who follow the headlines, but still feel like you're underusing AI at its full capability. Start with a 7-day free trial. Savi Security screens your family's texts, voicemails, and calls for realistic scams, including live-call monitoring during suspicious conversations, for $8/month. Muse Image helps you generate, edit, blend, and share images across Meta AI, Instagram, and WhatsApp, and also released a Muse Video preview. Claude for Open Source gives eligible open-source maintainers and contributors six months of Claude Max access—free for six months if eligible. Willow Frontier Mini gives you unlimited voice dictation for turning speech into clean text, while Frontier Pro targets faster and more accurate team workflows —free to try. Norm Ai helps legal and compliance teams use agentic law workflows for regulated work (raised $120M at a $1.2B valuation)—outcome-based pricing Antidoom gives model builders an open-source recipe for reducing reasoning doom loops with Final Token Preference Optimization.
| |
📰 Around the Horn |  | We literally don’t deserve these kings. If you currently pay for Wispr Flow, try this instead. |
| Illinois signed a state AI safety law that requires major developers to disclose safety practices, report incidents, and face independent audits starting January 1, 2027. OpenAI showed how computational astrophysicist Chi-kwan Chan is using AI to help build the first video of a black hole, after helping capture the first-ever black hole image with the Event Horizon Telescope. China reportedly discussed restricting overseas access to its most advanced AI models, potentially mirroring U.S.-style frontier AI controls. DeepSeek is reportedly developing its own inference chip to reduce reliance on Nvidia and Huawei (think I read Zai who makes GLM is doing the same?). Microsoft reportedly began replacing some OpenAI and Anthropic usage inside Excel and Outlook with its own MAI models to cut Copilot costs. Chinese-origin models have held more than 30% of U.S. OpenRouter token usage every week since February 8, with price doing a lot of the distribution work. Forterra deployed more than 100 autonomous Lancer vehicles in Ukraine, though many missions still need teleoperation. Samsung reportedly expects a 19-fold jump in quarterly operating profit as AI data-center memory demand stays hot. A Future of Life Institute assessment said major AI companies are weakening or dropping some voluntary safety commitments. Reddit said its AI-assisted defenses block 23M spam views per day and revoke nearly 2M fake votes daily.
| |
|
|
| Get the blueprint for modern privilege security. Break down the eight must-have PAM capabilities to secure human, machine, and AI identities — reducing attack paths and gaining visibility across Agentic AI, DevOps, OT, and Zero Trust. | Download the guide to strengthen your security posture. | |
📖 Midweek Wisdom |  | A screenshot of a tweet of a screenshot of a headline |
| Goodfire introduced Block-Sparse Featurizers (paper), a new way to inspect vision models by finding concept clusters instead of one blunt “feature” at a time. Allie K. Miller is getting good results with Fable by telling it to go out there and have some fun (us paraphrasing here) before getting into her actual goal. Nature published research on The AI Scientist, an agentic system that generated ideas, wrote code, ran experiments, analyzed results, drafted papers, and passed a first review round at a top ML workshop. Sakana AI released ShinkaEvolve, a system that uses language models to evolve new algorithms with far fewer search steps than traditional evolutionary methods. Lilian Weng argued that “harness engineering,” or the tools, verifiers, workflows, and scaffolding around models, may be one of the biggest unlocks for AI self-improvement. Lenny Rachitsky found tech workers splitting between AI-amplified and AI-destabilized, with burnout rising and career optimism falling below 50%.
| | |
| | That’s all for now. | | What'd you think of today's email? | |
|
| P.S: Before you go… have you subscribed to our YouTube Channel? If not, can you? |  | Going for an anime aesthetic this month! |
| P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here. |
|
