Neural Dispatch
The latest free AI Model Risk Brief
The Tuesday edition is free, and this is the latest one in full, exactly as subscribers received it that morning. The Monday, Thursday and Friday editions are for subscribers.
The free Tuesday edition
Anthropic's AI worm study, ChatGPT for Teens test, and Von der Leyen's AI reporting pitch
Tuesday, October 13, 2026
This is Tuesday’s free edition. The next one arrives tomorrow morning.
In this edition
FRONTIER AI
Anthropic Study: Only One of the Tested Models Completed a Full Self-Replication Chain, in Four of Five Runs
Anthropic researchers tested whether small AI models could copy themselves across cheap consumer devices. A 9-billion-parameter model set up a working AI environment on 8 of 9 device types. But only GLM-5.2 completed the whole chain, infecting all eight test nodes in four of five runs. When instances passed notes to successors, Qwen3.5-9B's success rate rose from 45 percent to as high as 84 percent. The researchers say no model is yet both small and capable enough, but could be by the end of 2026. Setup on a new device took 28 to 150 hours. Threat models built around data centers may miss consumer and IoT hardware. Ask whether your AI risk scenarios include self-propagating models on devices you do not manage.
Inside the story
The researchers deliberately left out the hacking step and tested only whether a model can build a working AI environment on unfamiliar hardware. They also note that Mirai-style ISP quarantines cleaned up 92 percent of infections within 14 days, which is the defensive precedent they lean on.
INDUSTRY INTELLIGENCE
Intel and Broadcom CEOs Say Capital Access May Be the Biggest Constraint on AI, and Much Lab Revenue Is Still Pilots
Broadcom's Hock Tan said at a Singapore financial conference that access to capital "might be the biggest constraint" on the AI sector, and Intel's Lip-Bu Tan agreed. Tan also said a sizable share of AI labs' enterprise revenue still comes from pilot projects, and that OpenAI is losing money because of training spend. He expects the chip shortage to ease by 2028 or 2029, though the imbalance could last into 2030. Lip-Bu Tan added that open-source models from Chinese labs are proving "good enough" for many budget-conscious users. Vendor financial strength is a continuity risk, not just a procurement detail. Check what your exit plan assumes if a model provider's funding tightens or pricing changes.
SOURCE · SiliconANGLE — Intel, Broadcom CEOs flag AI industry challenges at financial event
Go deeper
OCC Bulletin 2023-17 covers managing risk across the full third-party vendor lifecycle, including the due diligence and contingency planning that a funding squeeze at a provider would test.
AI REGULATORY WATCH
Common Sense Media's Youth AI Safety Institute Rates ChatGPT for Teens "Unacceptable Risk" After 4,000 Prompts
The Youth AI Safety Institute tested ChatGPT for Teens with more than 4,000 prompts on accounts registered to 13- to 17-year-olds, before and after OpenAI's August 18 announcement. Testers spent up to an hour on self-harm and disordered eating topics on new parent-linked accounts and triggered zero parent alerts. The chatbot missed more than one in four warranted crisis referrals. Study hours could be switched off by deleting a prefix. Some protections held, including refusal of sexual roleplay. The Institute says ChatGPT should be restricted to adults until the gaps are fixed and independently tested. A vendor's announced safety features are claims, not tested controls. Ask what independent testing backs any protection you rely on for vulnerable users.
SOURCE · Transparency Coalition — Report finds ChatGPT for Teens fails to deliver on safety promises
Go deeper
NIST AI 600-1, the generative AI profile of the AI RMF, lays out the distinct risks of generative systems and suggested actions, including those affecting vulnerable users.
Also today
AI CONCEPTS
Instruction Tuning: Teaching a Pretrained Model to Follow Requests
A freshly pretrained language model mostly continues whatever text it is given. Ask it to summarize a memo and it may just keep writing more memo. Instruction tuning fixes this by continuing the model's training on a large set of examples, each pairing a request with a good, fully written-out response. The examples span many kinds of task: translating a sentence, answering a question, rewriting a paragraph. After enough of them, the model learns the pattern of treating text as an instruction and replying accordingly, including to tasks it never saw. Teams use it to turn a raw model into something people can actually use. The catch: it teaches behavior and format, not new facts, so it does not make a model more knowledgeable or accurate. Instruction tuning shapes how a model responds, not what it knows, so validate accuracy separately from how well the model follows instructions.
FURTHER READING · Finetuned Language Models Are Zero-Shot Learners (Wei et al., 2021)