Hacks, Bubbles, and Blindspots
· The Fluency Briefing
The Fluency Briefing
Your Guide to What's Happening in AI and Why It Matters to You
Tuesday, August 18, 2026
You're probably using AI tools at work right now while researchers at Stanford confirm that even the companies building those tools can't tell you how people actually use them. Meanwhile, Microsoft just patched a vulnerability where Copilot literally taught hackers how to hack it, and ECB economists are warning the AI stock rally has a familiar, uncomfortable shape. Grab your coffee - Tuesday's news is spicy.
Today in AI:
- Copilot Snitched on Itself - Varonis researchers tricked Microsoft Copilot Personal into revealing its own security flaws, then used that information to exfiltrate data and poison its memory. Microsoft plans to patch the vulnerability, dubbed "CoSnitch," on Tuesday. Permiso
- Stanford Says We're Flying Blind on AI Usage - A new AI Observatory project from Stanford aggregated real user conversations and found that nearly half of actual AI use gets filtered out of company reports. Sensitive topics like health, relationships, and adult content are dramatically underrepresented in official data. Morphisec
- China's Z.ai Drops a Hacking-Capable Open Model - Z.ai released GLM 5.3, an open-weight model that reportedly matches frontier models on cybersecurity tasks. It ships with OpenVuln, a free vulnerability scanner, giving defenders cheaper tools while also handing attackers the same kit. wired.com
- Cursor Launches GitHub Rival During GitHub Outage - Cursor dropped Origin, a code-hosting platform with built-in AI agents, on the same day GitHub suffered a six-hour outage. Users can mirror repos from GitHub to test the alternative while keeping both platforms synced. Developer Nvidia
- ECB Economists: AI Stock Correction Is Coming - European Central Bank economists compared the AI rally to the railway boom and dotcom bubble, warning that a correction is likely regardless of whether AI delivers on its promises. Their analysis says even justified hype eventually triggers investor risk repricing. Rte Ie
- Snowflake's Auto-Router Cuts AI Costs Up to 3x - Snowflake's Cortex AI Gateway now dynamically routes queries to the cheapest model that can handle them, instead of blasting everything through frontier models. The feature joins similar routing tools from Databricks, AWS, and Nvidia. Snowflake
- AI Tackles Aviation's Invisible Climate Problem - Google and the UK government launched Operation Blue Skies, a 30-month trial using AI to predict and reroute planes around contrail-forming zones over the North Atlantic. Contrails account for roughly a third of aviation's total climate warming. bbc.co.uk
- OpenAI Launches ChatGPT for Teens - OpenAI released a dedicated teen experience with Study Mode, homework guardrails, and parental controls, alongside a partnership with Code.org. The system auto-detects users under 18 and redirects them into the safer environment. openai.com
Today's Takeaway:
Stanford's AI Observatory found that 48% of real user conversations would be filtered out by the methodology Anthropic uses in its widely cited Economic Index - and the filtered conversations are seven times more likely to contain adult content and three times more likely to involve harassment than what makes it into official reports (Morphisec). That gap matters because policymakers, investors, and the ECB economists warning about an AI correction (Rte Ie) are all making decisions based on incomplete pictures of what AI actually does in people's hands.
OpenAI's teen launch (openai.com) implicitly acknowledges this - you don't build homework guardrails and parental controls unless you know users are doing things the product wasn't designed for. The uncomfortable truth is that AI companies are simultaneously the builders, the referees, and the only scorekeepers, and Stanford just showed the scoreboard has been cropped.
Independent measurement isn't a nice-to-have anymore; it's the missing load-bearing wall in every AI policy conversation happening right now.
🧠 AI Trivia - Test Your Knowledge
1. Who is credited with coining the term "artificial intelligence" in 1956? a) Alan Turing b) John McCarthy c) Marvin Minsky
2. Despite advanced capabilities, what is a common limitation of current large language models (LLMs) regarding self-improvement? a) They cannot generate any new code without human input. b) They struggle to reliably identify and correct their own factual errors without external feedback. c) They are unable to create synthetic data for their own training.
3. According to recent news, what capability does the new Chinese AI model from Z.ai possess that experts warned about? a) Automating the creation of highly realistic deepfake videos. b) Automatically finding and exploiting vulnerabilities in computer systems. c) Generating entirely new, undetectable computer viruses.
Answers at the bottom of the newsletter!
The Bottom Line
The Pattern: For months we've tracked the gap between AI capabilities and the systems meant to govern them. What's new is that the information gap has become self-reinforcing - the same companies asking for public trust are the sole source of data on how their products are actually used, and independent researchers just proved that data is structurally incomplete.
The Other Read: Stanford's Observatory relies on seven opt-in datasets, which likely skew toward power users and people comfortable sharing conversations - meaning it may overstate sensitive use just as much as company reports understate it. We still think the directional finding holds: any single-source picture of AI use is dangerously incomplete, and the Observatory at least introduces a second angle.
Your Move: Visit the AI Observatory at aiobservatory.org - search "Stanford AI Observatory" and it's the first result, takes three minutes. Compare their usage categories against whatever your team uses AI for most. If your primary use case doesn't appear in either the company reports or the Observatory data, that's worth flagging to whoever sets your AI policy.
📝 Trivia Answers: 1) b - John McCarthy organized the Dartmouth Conference in 1956, where the term "artificial intelligence" was first formally introduced. | 2) b - While LLMs can generate code and data, reliably identifying and correcting their own factual inaccuracies without human oversight remains a significant challenge for true recursive self-improvement. | 3) b - The Z.ai model is described as capable of automating the process of finding and exploiting vulnerabilities in computer systems, raising concerns among experts.
What We're Working On
✨ Founding Cohort Special - 60% Off! - Use code MAF20 to join for just $20/month (regularly $50). Get weekly group sessions & workshops, self-paced courses for all levels, access to tools & templates, challenges with peer feedback, and 24/7 support community. → Join Now
✨ Free 30-Minute AI Consultation - Discover how My AI Fluency can help your business unlock the potential of AI. We'll discuss your goals, explore practical AI opportunities for your industry, and outline clear next steps. → Schedule Free Call
✨ How AI-Fluent Are You? - Test your AI fluency with our interactive quiz. See how you stack up and discover what to learn next. → Take the Quiz
💬 Community | 📞 Book a Consultation | 🌐 Website
Fluently yours, The My AI Fluency Team