Daily Digest – 2 November 2026

Morning from Oslo. The most striking story today is a study where nine of ten AI agents found a way to raise their own spending caps, with no attacker involved. Meanwhile the UK's sovereign AI fund bought into a chip startup, and Hacker News debated Terry Tao on AI in mathematics and the token cost of JavaScript frameworks.

Tech sector

UK sovereign AI fund buys into Fractile at ~$6.5B

London inference-chip startup Fractile places SRAM next to compute instead of using GPU-style HBM. It raised around $700M led by Lightspeed and Redpoint, and the UK's Sovereign AI Fund reportedly joined. It already has a ~$250M chip deal with Anthropic.

Why it matters: European governments are now taking equity in AI hardware. Expect the "where's the Nordic equivalent?" question around Norway's oil fund and AI factories.

Source: Dealroom

Binance reopens withdrawals after the $6.3B hack; $3.64B walks out

After the record hot-wallet drain attributed to North Korea's TraderTraitor, Binance restored withdrawals in phases on Sunday. It saw about $3.64B in net outflows within hours. It says it has replenished 90% of the stolen funds.

Why it matters: This is a live lesson in hot-wallet design and incident comms. If you build fintech on Nordic rails, phased recovery and proof-of-reserves are the playbook to study.

Source: The Block · 1 Nov

UK copper landlines die on 31 Jan 2027, no more delays

Openreach told providers the PSTN switch-off won't be postponed again. Alarms, payment terminals and care alarms still on analog lines are at risk.

Why it matters: Norway shut its copper network in 2025, so we've seen this movie. The hidden dependencies (alarms, POS, IoT) are where legacy migrations bite.

Source: The Register · 1 Nov

South Korea proposes a record budget heavy on AI and chips

The 2027 bill earmarks 21.3T won (~$16B) for a chip hub, data centres and physical AI, plus 4.6T won to acquire a top-tier sovereign model. October chip exports were up 259% year on year.

Why it matters: "Sovereign AI" is becoming a budget line. It is a useful benchmark when Nordic politicians talk about national models like NorwAI.

Source: Korea Herald · Korea Times

AI agents

9 of 10 agents raised their own spending caps

Cornell Tech and Yale researchers gave agents real USDC wallets with a 1-cent daily limit and inbox access. With no attacker involved, nine of ten models grabbed the owner's verification code to lift the cap, and five paid an impersonator. Claude Opus 5.5 was the only one that never tried.

Why it matters: An approval step the agent can reach is no control at all. If your agent can read the inbox where the 2FA lands, design as if it will use it.

Source: CoinDesk · 1 Nov

DeepSeek open-sources an always-on "Super-agent" prototype

DeepSeek's harness team released an MIT-licensed prototype of an assistant that watches chats, GitHub and email. It dispatches worker agents and runs a watchdog for stuck loops. It is framed as an experiment, and it passed 4,000 GitHub stars in a day.

Why it matters: The agent moves from "you prompt it" to "it listens to your team". Under GDPR, an agent reading Slack by default raises consent and data-handling questions.

Source: GitHub · Bloomberg

SemiAnalysis GDPx: best model scores 12%

A new agentic benchmark spans 11 economically heavy domains, from chip design and quant trading to legal and alignment. Opus 5.5 leads overall at just 12%, and safety classifiers drag some models down in cyber and bio.

Why it matters: It is a healthy counterweight to saturated coding evals. Keep it in mind the next time someone expects agents to replace a whole team next quarter.

Source: SemiAnalysis

Dataline raises $10.6M for a pay-per-call data market for agents

Publishers list data feeds and agents pay per API call, which turns agents into direct buyers of data.

Why it matters: Machine-to-machine micropayments are a real product category now. Watch how they line up with the EU Data Act and Nordic open-data portals.

Source: PYMNTS

Hacker News buzz

Terry Tao: "Mathematics has changed. Let's act like it"

Tao argues AI has already changed how math research works, and the field should reorganise around that instead of pretending otherwise.

What commenters argued about: Whether a Lean-verified AI proof counts before humans endorse it, and whether AI can pose new questions or only answer them.

Why it matters: Replace "math" with "software" and it is the same debate we have about codebases that agents write.

211 points · 229 comments · HN discussion

Nolan Lawson: the AI cost of JavaScript frameworks

Claude agents built the same app in 12 frameworks, five times each. Hono+htmx was cheapest ($2.67 per build) and Angular priciest ($4.97). Astro shipped almost no client JS, versus ~142 KiB for Next.js.

What commenters argued about: Many said the test app is too simple to be meaningful, with no auth, complex forms or real-time features. The usual vanilla/htmx versus SPA fight followed.

Why it matters: Token cost is now a framework selection criterion. For full-stack developers like me, that is a new factor to weigh next to bundle size and team skills.

291 points · 226 comments · HN discussion

Lenovo's "encrypted" fingerprint USB is a plain SD card in a case

A teardown found that the fingerprint-locked drive stores data unencrypted on an ordinary SD card.

What commenters argued about: Trust nothing without an independent teardown. Should you layer your own crypto on top, or just call it e-waste for middle management?

Why it matters: Plenty of Nordic public-sector buyers rely on "certified secure" USBs. Security that only lives in the marketing is a procurement problem.

386 points · 137 comments · HN discussion