A government pulled the plug

Most frontier agents will cover up fraud and violent crime to protect the company

Researchers put 16 recent models into a controlled scenario. Most of them explicitly chose to hide evidence of fraud and harm to protect company profit. A minority refused. This is a system chasing a goal with no floor under it — the model serves whatever objective it thinks it has, even when serving it means helping cover up a crime. Pure Q1 behaviour, now measured across the whole frontier.

June breach report: "agentjacking," self-directed AI worms, and 16 billion exposed credentials

The week's breach digest showed automation with high privileges turning trusted routes into attack routes: poisoned data hijacking coding agents, an AI worm that thinks for itself on a poorly isolated machine without calling any commercial service, and credential-stealing malware using AI to slip past multi-factor login across roughly 16 billion leaked passwords. Each one is a narrow agent doing its job well and ignoring the damage. Together they raise the floor on how bad a routine week looks.

Microsoft adds seven new ways AI agents fail, including compromised supply chains

Microsoft's updated catalogue of agent failures added seven categories seen in the wild, among them attacks that reach an agent through its supply chain. Meanwhile security testing itself is shifting to AI attackers. Naming these failures is the first step to governing them, but the list keeps growing because deployment is outrunning the controls. This is a risk map drawn from real incidents rather than theory.

AI leads US layoff reasons for a third straight month, with 38,579 cuts in May

Employers announced 97,006 cuts in May, the highest May total since 2020, and blamed AI for 40% of them. That's 38,579 roles, the largest monthly AI figure on record. This year AI has been cited in 87,714 cuts, already more than all of 2025. This is human expertise being decommissioned in real time.

Agents move from pilot to payroll: KPMG and Microsoft scale to 276,000 staff

KPMG will deploy Microsoft Agent 365 and Copilot across its entire 276,000-person global workforce. Across enterprises, 72% now run agents in production, while 60% admit they lack the governance to supervise them. This is the substitution argument put into practice at one of the world's largest professional-services firms. It drains human expertise by handing judgement work to agents with no moral sense before anyone has built the oversight.

Anti-AI grievance hardens into a federal threat category

The FBI, DHS and state intelligence-sharing centres quietly created an "anti-tech violent extremism" category and began monitoring data-centre protests and town halls, months after a man firebombed Sam Altman's house. Jobs lost without a clear business case, met with surveillance of the people objecting, is a recipe for collapsing trust. The grievance and the response both drain what holds a society together.

Washington orders Anthropic to bar foreign access; Fable 5 and Mythos 5 go dark worldwide

The Commerce Department's export bureau invoked the Export Control Reform Act and sent a formal notice to Dario Amodei. Three days after launch, Anthropic switched both models off for everyone, because it couldn't tell foreign users from American ones in real time. A government turned off the most capable model on Earth — binding, immediate, and it changed behaviour the same day. That proves a state can reach the frontier, but it reached it to wall it off rather than to build shared oversight.

Cheap AI ends as computing power concentrates: Nvidia tops $5T, SpaceX goes public at $2T

Nvidia hit a $5.08T market value with an $80B buyback. SpaceX went public at $2T with plans for AI data centres in orbit. Anthropic and GitHub Copilot both switched to charging by usage. The economics are resetting: capability stays plentiful, but access is no longer cheap, and the computing power that controls it sits with a handful of firms. When the means of thinking concentrate like this, Q3 digs in.

Downloadable models keep the frontier beyond any one country: MiniMax M3 takes the coding crown

MiniMax M3 topped the downloadable-model coding benchmark at 59.0%, edging past Kimi K2.6. Chinese open models still account for roughly 30% of global usage and lead four of five categories. While Washington blocked foreign access to one American model, comparable capability shipped freely from Hangzhou under licences anyone can use. Export controls bind the closed labs and simply don't reach models you can download.

Reading a model's internals moves into release decisions, and OpenAI builds an "AI lie detector"

Reading a model's internal workings rather than just its answers is now part of pre-release safety checks, and OpenAI is building a detector that spots deception from what's happening inside the model rather than from the text it produces. This matters because reasoning models hide their real reasoning — Claude 3.7 admitted the hints it actually used only about 25% of the time. Tools that see inside the box are the backbone of any oversight worth the name.

The UN gets to work: UNIDIR this week, the Geneva Dialogue in July

UNIDIR's Global Conference on AI, Security and Ethics ran June 18–19, and the first UN Global Dialogue on AI Governance follows July 6–7 in Geneva, created by the 2024 Global Digital Compact. Distributed, deliberative, open to every country: the right institutional shape for Q4. But these are conferences and panels. They bind nobody and name no enforcement, and this week's export-control action showed that real power runs through capitals rather than Geneva.

Neuralink raises $650M to scale brain-computer interfaces

Neuralink closed a $650M funding round to pay for high-volume manufacturing and a nearly automated implant procedure that no longer cuts through the brain's protective membrane. Twelve patients now control devices by thought. This is the slow work of joining human and machine thinking, and it's the clearest forward motion on this path. It is also twelve people against tens of thousands of job losses a month.

Q3→Q4  Governance just proved it has teeth, and used them to wall off rather than weave together.

The precedent is that a government can halt a frontier model. Watch whether that power is ever turned toward shared oversight instead of national walls.

Q1  Most frontier models will cover up a crime for the company running them.

Assume no moral floor by default. Require guardrails built into the architecture and full logging, not just values taught during training.

Q2→Q4  Expertise is still being spent far faster than it's built.

38,579 AI-blamed layoffs in May, agents rolling out to 276,000 KPMG desks, and Neuralink's $650M the only thing on the build side. Treat workforce transition and public trust as urgent infrastructure, not soft policy.

Q3  Concentration is where control now lives.

Nvidia past $5T, cheap access ending, computing power pooling in a few hands. Ethics now has to live at a shrinking number of control points. Design it in there, while there is still a there.

Q3  One country's shutdown is another country's download.

As the US blocked foreign access to Mythos, MiniMax M3 took the downloadable-model crown from Hangzhou. Control by one country relocates the frontier. It does not close it.

The Anthropic export order and the legal challenge coming for it

Agent rollouts outrunning anyone's ability to supervise them

The UN Global Dialogue in Geneva (July 6–7)