Tech news, 31 October 2026: Binance's record hack, Intel's guidance shock, open weights at the top end, OpenAI's new Model Spec, voice AI consolidation and New York's chatbot law

1. Binance reports a $6.3 billion theft, the largest on record

Binance said on 29 October that attackers had stolen crypto assets it estimated at about $6.3 billion, suspended deposits and withdrawals, and said it would cover losses². Its early account is that a compromised laptop gave the attacker access to a service container, from which unauthorised withdrawals were submitted; markets fell and stablecoin issuers froze part of the funds¹. Figures were still moving at the time of writing.

CTO take: the pattern applies to any platform that moves money: one compromised internal service should never be enough to release funds. Check that services can only request payments, that an independent policy layer approves them, and that limits cap what can move in an hour. The full analysis is in the Binance hack: architecture lessons.

2. Intel's weak outlook sends its shares down sharply

Intel's third-quarter results beat estimates, but its fourth-quarter revenue guidance fell well short, and CNBC reported a drop of about 21% in the shares as the company pointed to a PC slowdown lasting into next year³.

CTO take: for buyers, this is a reminder that chip supply chains are under strain from several directions at once. If your hardware roadmap depends on a single vendor's next node or product, keep an alternative in the plan, especially for edge and on-premises AI.

3. Mistral opens its trillion-parameter model; Canada backs Cohere

Mistral released the weights of Mistral Large 4, a one-trillion-parameter mixture-of-experts model aimed at security teams who want to run it themselves; the benchmarks quoted are Mistral's own⁴. The same day, Canada announced an equity investment of up to C$1.6 billion in Cohere, tied to building compute capacity in Canada⁵.

CTO take: sovereign and self-hosted options at the top end are real now, but holding a trillion parameters in memory is expensive. Test through an API first and keep the stack portable. See should you run a 1T model yourself? and a sovereign LLM platform for a European telecom.

4. OpenAI rewrites its Model Spec

OpenAI published an updated Model Spec on 29 October, organised around five core values and an explicit order of authority between system, developer and user instructions; it describes the document as a target its models do not yet fully meet⁶.

CTO take: if you build on OpenAI models, the instruction hierarchy is the part to read: it affects how your system prompts interact with user input and tool output. Re-run your evaluation and prompt-injection tests after model updates rather than assuming behaviour holds; see prompt-injection defence for agents.

5. Deepgram buys Daily and sets up a Pipecat Foundation

Deepgram announced it is acquiring Daily, the company behind the open-source Pipecat voice agent framework⁷. Both companies say Pipecat stays under its existing BSD licence and will be governed by a new independent foundation with a steering committee that includes Deepgram's competitors⁸.

CTO take: voice agents are consolidating around a few stacks. If you depend on Pipecat, the foundation is good news, but track its governance and keep your speech-to-text and text-to-speech providers swappable behind your own interfaces.

6. New York signs a youth chatbot safety law

New York's governor signed a law on 28 October barring a set of AI chatbot features judged unsafe for minors, including chatbots that pretend to be human or encourage self-harm⁹.

CTO take: any public chatbot that New York users can reach should plan for age-aware behaviour, enforced guardrails and evidence. See what AI product teams must build and the US page.

What I'd do this week

• Run an assume-breach review of every path from an internal service to an action that moves money.
• Re-run evaluations and prompt-injection tests on any model that changed this month.
• Check which open-source frameworks in your AI stack have changed owners, and who governs them now.
• If you serve consumers in New York, map every chat surface against the new law.

Working across these markets? See where I work, or analyse your challenge and bring the result to a call.