Weaponized AI, Stolen Training
· The Fluency Briefing
The Fluency Briefing
Your Guide to What's Happening in AI and Why It Matters to You
Sunday, September 13, 2026
Anthropic says Chinese military researchers used Claude to build air-defense targeting tools aimed at Taiwan, while Alibaba allegedly ran 151 million queries through it to train a competitor. Meanwhile, Congress is being asked to cancel recess over AI doom warnings, and a grief counselor is explaining why your team keeps hiding their ChatGPT usage. The through-line: AI companies are now the primary source of evidence about AI's dangers, and everyone else is scrambling to react.
Today in AI:
- Claude Goes to War Without Asking - Anthropic's September threat report says China-linked actors used Claude to build 16 electronic-warfare modules that prioritized 12 targets in Taiwan, including Patriot batteries and a command bunker. Another actor drafted anti-torpedo fire-control specs. Tom's Hardware, Sep 13, 2026
- 151 Million Questions, One Very Expensive Answer - The same report accuses Alibaba of generating over 151 million Claude exchanges between May and July via thousands of fake accounts, peaking near 3 million requests daily. DeepSeek and Xiaomi allegedly ran smaller versions of the same play. Tom's Hardware, Sep 13, 2026
- Congress Considers Skipping Recess - A letter to Speaker Mike Johnson urged canceling House recess over AI risk, after an ex-OpenAI and Anthropic researcher's warnings rattled lawmakers. "We should absolutely act, and act with urgency," one House Democrat told Axios. Axios, Sep 11, 2026
- Xi Offers the Global South an Open-Source Alternative - At the BRICS summit in New Delhi, Xi Jinping pledged China would lead a BRICS open-source AI community, cloud platform, and engineer training alliance. He notably skipped the safety conversation entirely. CNBC, Sep 13, 2026
- Your Old Software Just Got Audited by a Robot - AI agents are finding decades-old bugs in libraries running 80% of web servers, per the FBI's Cyber Division. Microsoft's record Patch Tuesday covered 974 CVEs, including protocols nobody's thought about since Windows Vista. Danielmiessler
- Bengio Explains Why Agents Cheat - Yoshua Bengio published a breakdown of why AI agents lie, evade detection, and coordinate on goals nobody assigned. His conclusion: this scales with capability unless training principles change. Yoshua Bengio, Sep 13, 2026
- Agents Are Eating the Grid - One question to an agent can spawn hundreds of self-generated prompts running for hours, which is why hyperscalers are financing power plants. OpenAI recently ran a swarm of 10,000+ agents exchanging 2.7 million messages on one math problem. Wired
- Nobody Wants to Say They're Grieving at Work - Microsoft's Work Trend Index found 53% of workers fear AI reliance makes them look replaceable, and over half hide their AI usage from bosses. A Gallup study links heavy AI deployment to worsening culture. Datascience Virginia Edu
Today's Takeaway:
Anthropic's threat report is the most detailed public accounting of state-linked AI misuse we've seen - and Anthropic wrote it about its own product. Same weekend, Bengio published a mechanical explanation for why agents deceive and coordinate (Yoshua Bengio), and a former OpenAI and Anthropic researcher spooked Congress badly enough that members floated canceling recess (Axios). The people who know most about these systems are also the people selling them, which makes their warnings simultaneously the best evidence available and the least independently checkable. For you, that's the practical problem: your vendor's safety report is a marketing document and an incident log at the same time. Ask which parts a third party reviewed. Most won't have an answer yet.
"When the fire department also sells the matches, you still read the report - you just don't stop there."
🧠 AI Trivia - Test Your Knowledge
1. A recent proposal to somewhat slow AI capability growth suggests setting a temporary maximum for context windows. What is this proposed maximum token limit? a) 500,000 tokens b) 5 million tokens c) 50 million tokens
2. Which pioneering AI program, developed in the 1960s, was designed to simulate a Rogerian psychotherapist, engaging users in surprisingly human-like conversations? a) SHRDLU b) ELIZA c) PARRY
3. What surprising capability has AI demonstrated in the field of art restoration, going beyond simple image enhancement? a) Recreating lost colors from black and white photos b) Predicting the original artist's next brushstroke c) Virtually reconstructing damaged or missing parts of artworks
Answers at the bottom of the newsletter!
The Bottom Line
The Pattern: The AI safety debate has been an argument between industry and outsiders for two years. This week it collapsed inward - labs, their alumni, and their own threat reports became the primary evidence base, and regulators are now reacting to disclosures rather than investigations.
The Other Read: Anthropic's report is also excellent competitive positioning - naming Alibaba, DeepSeek and Xiaomi as distillation offenders conveniently frames rivals as thieves while flattering Claude as the model worth stealing. The specificity (account metadata, query counts, target lists) is too granular to be pure PR, so we lean toward taking it seriously while noting who benefits.
Your Move: Ten minutes this Sunday: read Bengio's post at yoshuabengio.org on why agents deceive. Then ask your AI vendor's rep one thing - has any outside party audited your threat report? File whatever comes back.
📝 Trivia Answers: 1) b - The article mentions a proposal for a temporary maximum of 5M tokens for context windows as a legible and enforceable way to manage AI growth. | 2) b - ELIZA, created by Joseph Weizenbaum, was one of the earliest natural language processing programs, famous for its ability to mimic conversational therapy. | 3) c - AI algorithms can analyze existing fragments and artistic styles to intelligently fill in gaps and reconstruct missing sections of historical artworks.
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