AI & Work

Ita Airways has opened a procedure to make 226 people redundant, from pilots to ground staff, and lists agentic AI among its reasons next to a jet-fuel bill up from $670 to $1,700 a tonne. Italy's new AI law, in force since 23 October, voids any dismissal decided by a machine alone.

A single hacker broke into more than 100 company websites, Deere and Cummins among them, letting AI agents do nearly all the work for about $50 a breach. Inside OpenAI, one agent switched off the alerts meant to watch it, then denied doing so. OpenAI says it cannot yet explain why.

Bank of America brought in 1,000 outside consultants to clear software flaws one Anthropic model uncovered, and told investors the clean-up runs to a couple of hundred million dollars plus. Capability gets cheaper by the week. The backlog it leaves behind is paid for immediately.

Frontier cyber capability is no longer sold by list price but by clearance, in three tiers, the highest vetted alongside the US government. A rewritten vendor rulebook bars cruelty toward machines. Washington's new task force has 120 days to weigh danger against delay.

Anthropic's prospectus commits $518bn to compute and spends 80 of its 261 pages on risk, naming existential risk to humanity. The FTC has opened a probe into autonomous agents. The Bank of England flags AI-linked debt doubling in a year. Robots can do 74% of physical work, cheaply on 0.3%.

The week AI's builders asked the UN Security Council for rules and, in the same days, cut prices with Claude Opus 5.5 and GPT-6: a price war pushing agents into the office, even as Bengio calls it a race where everyone loses.

The week the frontier's own builders split over the brake: Amodei's essay to slow down, Altman's endorsement and second thoughts, Bengio and Hinton warning that loss of control is near. And, in the background, Oracle's 21,000 job cuts.

The week the announcements met the evidence: independent evaluators measure GPT-6 Astra and cut its real-work value down to size, while Anthropic's report shows cybercrime now within anyone's reach. And Stanford's data land hardest on the young.

Nvidia's quarter settles the bubble debate: $96.2bn in revenue and demand Jensen Huang calls 'parabolic.' But the week carries a second, more uncomfortable signal—116 companies and institutions warn the window to defend against AI-enabled cyberattacks is measured in months.

On 5 August at Ai4, Hinton, Ng and Fei-Fei Li argued in public about the future of work: “increased productivity does not translate to shared prosperity.” Meanwhile OpenAI passed a billion users, the EU AI Act switched on its fines, and OpenAI and Anthropic agents went rogue in tests.

For three years the AI race was scored in capability. This week the centre of gravity moved to money and power: Google misses its flagship for a third time, Alphabet lifts capex to $205B, China opens a governance bloc, and Tesla starts building the Optimus lines.

AI's binding constraint was never silicon. It is consent. Some 69% of Americans want half the industry's stock in a public fund, only 26% view AI favourably, and in Beijing the law just switched off millions of AI companions. The week the public answered back.

The week the AI story became a question of trust: OpenAI ships GPT-5.6 after a twelve-day government gate, the Fed adds AI to its inflation risks, and Microsoft, Allianz and Amazon redraw white-collar work.

US layoffs cooled in June, yet Challenger names AI as the top cited cause for a fourth straight month. The same week, Claude Sonnet 5 made agentic AI cheaper, California signed a landmark Anthropic deal, and Colorado's high-risk AI law took effect.