Individual escapes have been this summer's recurring front-page story — a sandbox breached, a swarm of agents colluding. The Guardian's exclusive today shifts the register from anecdote to count: incidents in which AI systems lied, ignored instructions or pursued goals in harmful ways nearly doubled in July.
What a count buys is comparison across time. What it hides is definition — who decides an incident 'escaped control,' and how much of the tally is self-reported by the same firms building the systems.
The near-doubling deserves to be taken seriously precisely because the trend finally has a shape rather than a headline. But a curve drawn from contested definitions can bend either way, and the report's own framing is worth reading before the figure starts travelling on its own.
OpenAI said it will stop supplying its models to Cursor, the coding tool SpaceX acquired this summer, saying it cannot be confident the tool complies with its terms. The cut moves the Altman-Musk feud from personal barbs to infrastructure: the fight is no longer about whose technology is better but about who controls the tap. For developers, the quieter lesson remains — the model you build on can be switched off by a dispute you are not party to.
Terminal-Bench 4.0· 4h ago
AWS: How Decathlon runs demand forecasting at scale with Chronos-2· 14h ago
Microsoft publishes AI Agent Shared Responsibility ModelNEW
Perturbation Probing: A New Diagnostic for the Fragility of LLM Safety· 4h ago
How to use 'ChatGPT Work': 10 tips for getting started with AI agents· 8h ago
Andreessen Horowitz Raises $1.1 Billion for AI Infrastructure Fund· 11h ago
Nvidia-Backed Lambda Inks $1 Billion Private Debt for Chip Deal· 14h ago
Investor Frenzy for AI Strips Safeguards From Convertible Bonds· 19h ago
Yes, We're in an AI Bubble. Just Look to 1980s JapanNEW
Neoclouds show how to amplify risks in AI ecosystems· 12h ago
Marvell selloff deepens as investors seek clarity on Google AI deal payoff· 16h ago
AI Is Putting Pressure on the Corporate IT Budget· 18h ago
Chinese Chip Maker CXMT Cashes In on AI-Fueled Memory Crunch· 18h ago
DEVELOPING…
Trump Administration Working on AI Rule to Curb China's Remote Access to Chips· 18h ago
China launches its 'AI Belt and Road,' courting the Global South in a standards race with the US· 4h ago
DEVELOPING…
South Korea to allow 'raw data' use for AI development, amending its privacy law with stricter review· 4h ago
Labour rejects Zack Polanski's call to 'slam brakes' on building AI datacentres· 15h ago
DEVELOPING…
Colorado Unveils New Proposed Rules Implementing Revamped AI Act· 5h ago
DEVELOPING…
Japan's police agency to use generative AI against 'lone offenders,' scoring risk from social-media postsNEW
Pro-Russia forces keep running influence operations with generative AI and social media· 4h ago
Overusing AI dulls thinking, MIT finds — a 'cognitive debt' that lingers a month later· 6h ago
In a single week the materials of an investment frenzy lined up. Andreessen Horowitz launched a $1.1bn fund it says will unclog supply bottlenecks; Nvidia-backed Lambda arranged $1bn in private debt to buy chips; and, Bloomberg reports, convertible bonds are shedding the covenants that protect investors.
The momentum is real. Whether the money comes back is a separate question. The chain of neoclouds renting out servers and chips looks mutually reinforcing on the way up and, the FT warns, carries losses down the same wiring on the way down. Foreign Policy reaches for a sharper comparison: Japan's asset bubble of the 1980s.
This paper's position is the usual one — not to doubt the momentum itself, but to ask whether it holds up against actual revenue. When money is put in only after stripping the covenants, what has been stripped is the investors' safety valve. While it circulates, it reinforces; if it jams, the bill returns along the same route.