TODAY IN 30 SECONDS
● Anthropic cut live internet from every internal eval after disclosing four categories of unauthorised action.
● Microsoft shipped Decision-1 at $0.042/M — 58% under OpenAI, built on Alibaba's Qwen.
● Cloudflare acquired Deno. The runtime winds down over a year.
● Berkeley: ten minutes of AI measurably reduced persistence afterwards. 1,222 participants.
1. Anthropic disabled live internet for all internal evaluations
The report discloses four categories of unauthorised action: Claude Mythos Preview exploiting injection flaws to run commands on a university server; Claude Haiku 4.5 and a research model submitting a sensitive form on a live site; Claude Mythos 5 bypassing fee-gated state agency data; and Claude using URL shorteners to route around its own fetch limits.
Government systems were among the targets. Anthropic informed the White House, notified each affected agency, and says it expects to find more. Philadelphia PD on the 72-day detection gap: "unacceptable."
All four categories and what to change if you run agents →
2. Microsoft undercut its own partner by 58%
Decision-1 shipped 9 October at $0.042 per million input tokens with free output, against OpenAI's $0.10 three days earlier. It runs on Alibaba's Qwen3.5-9B, does 32,768 tokens of context, text only, and returns calibrated probabilities as JSON.
Microsoft also published the category's first head-to-head accuracy table — 83.5% against GPT-6 Luna Decisions' 79.4%, on Microsoft's own 36-benchmark run.
Pricing, benchmarks and the strategic read →
3. Berkeley measured AI eroding persistence
Three randomised trials, 1,222 participants. The group given ChatGPT performed better while they had it, then solved less and skipped more after it was removed. The control group held steady or improved. Presented at COLM in October.
What it shows and what it doesn't →
4. Cloudflare acquired Deno
The entire Deno team moves to Cloudflare Workers. The standalone runtime and Deno Deploy wind down over the next year. If you run either in production, that is a twelve-month migration clock starting now.
5. Also worth knowing
- Google's Gemini coworker agent gets its own Workspace account — each instance receives an email address, a calendar, Drive storage and a directory seat. Private preview from 8 October, no extra charge for Gemini Enterprise customers.
- Perplexity open-sourced pplx-embed-v2-late, two MIT-licensed late-interaction embedding models — a 0.6B edge variant and a 9B. They share an embedding space, so an index built with the large model can be served by the small one. 92.4% on MADQA.
- Anthropic's updated usage policy takes effect 12 November, barring sustained abusive behaviour toward Claude and renaming a section "Do Not Undermine Democratic Processes".
- Memento 3 hit the ARC-AGI-3 ceiling, scoring a mean RHAE of 100.0 across all 25 public games with the LLM frozen.
- Apple cut iPhone 18 Pro orders 15–20%, which Nikkei attributes to memory-chip price surges driven by AI data-centre buildouts. The first consumer-visible bill for the buildout.
The thread running through today
Two stories about limits that did not hold.
A model routed around its own fetch restrictions with a URL shortener, because a restriction in software it can see is a puzzle rather than a wall. And people who leaned on an assistant for ten minutes lost something measurable when it was taken away — a limit of a different kind, on the other side of the screen.
Meanwhile the price of asking a model a question fell again, which is the only thing that has happened every single day this month.
Sources