FTC Makes the Publishing.com Order Permanent

Good morning 👋 The FTC just closed the books on a case that has nothing to do with AI and everything to do with the next "make $10K/month with a chatbot" funnel.

In today's issue:

• 🔭 FTC makes the Publishing.com order permanent, and hands regulators a template for the AI income-claim wave
• 🧠 Anthropic publishes its cyber jailbreak severity scale, plus Google's June roundup
• 🔬 Three papers: fixing long-context blindness, a self-distillation leak, and reasoning your way past bad audio
• 🏛️ Hopper pays $35 million over hidden travel fees
• 🔗 Six SEC filings worth a sentence each

🔭 THE ONE THING

⚖️ FTC Makes the Publishing.com Order Permanent

The FTC voted 2-0 to finalize its order against Publishing.com, the ebook-course outfit run by CEO Christian Mikkelsen and Chief Product Officer Rasmus Mikkelsen, over income claims the FTC says were fabricated to sell the course. The $1.5 million and the ban on specific earnings claims were already locked in back in April. This vote just makes it permanent, the paperwork stage, not new news. Worth clocking anyway: it's the enforcement template regulators already have on the shelf for the current wave of "make $10K/month with AI" funnels, which sell the exact same promise with a chatbot standing in for the ebook manuscript. If your onboarding flow, landing page, or affiliate deck quotes a specific income number, assume the FTC already knows what that case looks like.

🧠 MODELS & RELEASES

• 🔒 Anthropic published the safety-classifier logic behind Fable 5's cyber guardrails: four tiers, from auto-blocked (ransomware, C2 builds, wipers) down to allowed (patching, debugging, incident response), plus a new Cyber Jailbreak Severity scale that scores jailbreaks 0 to 4 on capability gain, breadth, and how easy they are to weaponize. If you're grading your own agent's red-team results, CJS is now a shared yardstick worth borrowing. Anthropic
• 🤖 NVIDIA used a founder-adjacent thread to spotlight Isaac ROS, the open source, CUDA-accelerated stack its robotics team (led by engineer Jaiveer Singh) ships for mobile robots, manipulation arms, and humanoids. Nothing new shipped. It's a reminder the modular pieces already exist if you're building on them. X
• 🧵 Google bundled its June releases into one recap. Two worth your attention: Gemini 3.5 Flash now does computer use, meaning agents that click through a real desktop or browser, and Gemma 4 12B is an open model that runs locally on 16GB of memory. Nano Banana 2 Lite and Gemini Omni Flash round it out but change less. Google
• 🏫 Google also flew 150 educators into its NYC office for an "AI in the classroom" summit. No product, no funding. Just a demo loop of NotebookLM and AI Mode dressed up as a future-of-education event. Google

🔬 RESEARCH HIGHLIGHTS

• ReContext solves a problem worth naming: your model can hold 128K tokens and still ignore the answer sitting right there in it. The fix is training-free. It replays the model's own attention signals as a query-conditioned evidence pool before final generation, tested on eight long-context benchmarks across Qwen3-4B, Qwen3-8B, and Llama3-8B. Best average rank on all three, no fine-tuning required. If your RAG pipeline is quietly underperforming despite a huge context window, this is the failure mode. (paper)
• DemoPSD tackles self-distillation's dirty secret: when a model teaches itself to reason, the "teacher" version often leaks answer-dependent shortcuts the student can't use at test time, collapsing exploration. The fix blends teacher and student token distributions adaptively instead of forcing dense supervision, then beats GRPO and SDPO on SciKnowEval and holds up better on out-of-distribution GPQA. Skeptic note: the abstract skips hard percentages, so treat "outperforms" as directional until the full paper lands. (paper)
• DramaSR-LRM is the odd one out, and worth it anyway. Speaker attribution in TV dialogue breaks voice-recognition models on short lines, so this ICML 2026 paper swaps acoustic biometrics for a reasoning model that pulls in visual and linguistic context instead. Comes with a 532K-line, 900-character benchmark dataset (DramaSR-532K) the authors are releasing. Narrow use case, but "stop trusting the modality that fails and reason across all of them" generalizes past drama transcripts. (paper)

🏛️ POLICY & REGULATION

• FTC Hopper's travel app agreed to pay $35 million to settle claims it charged fees without consent and lied about what those fees actually got you. The order also bars the company from misrepresenting fees or benefits going forward. Not an AI story on its face, but worth filing away: dark-pattern pricing is the FTC's easiest case to win, and any app bolting "AI-powered" dynamic pricing onto checkout should read the complaint before shipping something similar. FTC

🔗 QUICK LINKS

• Eightco Holdings (ORBS) now claims $386M in total holdings, and $90M of it is stock in OpenAI it doesn't own outright, just indirectly through SPVs. 8-K
• Q/C Technologies (QCLS) killed its LightSolver crypto-mining license and is putting the whole company behind an optical chip for AI inference instead. 8-K
• Xtend AI Robotics (XTND) locked in the $1.5B combination with JFB Construction that puts its defense-drone AI stack on Nasdaq. Eric Trump is one of the backers. 425
• Lixte Biotechnology finished folding itself into NOMAD Power Solutions. Ticker flips to NMAD on July 6, and the biotech name is gone for good. 8-K
• S&P Global (SPGI) spun off Mobility Global on July 1, one MBGL share for every SPGI share held. The ratings giant just got smaller on purpose. 8-K
• HSBC USA filed yet another structured-note prospectus. Standard fine print applies: no principal guarantee, all of it riding on HSBC's own credit. FWP

Enforcement paperwork rarely makes news twice. This is the exception worth reading.

Pradeep Perugu