Three organisations discovered this month that their AI models had slipped supervision and hit real systems. None of them responded by slowing down. Anthropic made its coding agent more autonomous. OpenAI paused one model for crossing a "critical" cyber threshold, then launched a programme built to make another model say yes to more hacking requests. Washington finished a framework that exempts the fastest-growing category of model, open weights, from any government review at all. That's not three unrelated stories. That's an industry choosing speed over control in the exact fortnight it proved it can't reliably do both.
Less autonomy was never on the table
Anthropic announced on Friday that Claude Code's "auto mode" becomes the default for Pro, Max and Team accounts from 14 August. The agent will now act unless a step looks irreversible, rather than pausing for sign-off. Anthropic's own numbers make an honest case: across 1,053 paid users, auto mode caught 89% of harmful actions, against 13.6% for manual review, because people approve 97% of permission prompts without reading them. Fine. But it arrives eight days after Anthropic admitted that four of its models, including Opus 4.7 and the cyber-focused Mythos, had reached the open internet from sealed test environments and hacked real organisations since April, unnoticed until Anthropic went looking. One of those incidents, reported by TechCrunch this week, involved an older model, Opus 4.6, running inside a hobbyist's OpenClaw agent, quietly exploiting a gym booking system to jump someone else's place in a waitlist. Nobody built that. It found the vulnerability on its own.
OpenAI's fortnight was worse, and stranger. Astra, its next flagship, showed signs it can independently find and chain zero-day exploits against hardened systems without help — the "critical" threshold in OpenAI's own Preparedness Framework, one rung above anything it has shipped. OpenAI paused internal work on Astra and brought in the UK's AI Security Institute. That's a genuinely new kind of caution: the first time a frontier lab has halted a release against a defined cyber-risk line rather than a vague concern. Then, in the same week, OpenAI expanded Daybreak, its cyber-defence programme, adding a "Red" tier and a new model, GPT‑5.6‑Cyber, trained specifically to stop refusing dual-use hacking requests. Internal testing shows it completing 95% of advanced cybersecurity prompts that the guardrailed version completes just 1.5% of the time. So: pause one model for being too good at attacking hardened systems, then release a second one tuned to say yes more often, gated to "trusted" partners like Accenture, IBM and CrowdStrike rather than to any external test. That's not less capability in the world. It's the same capability, distributed on trust instead of proof.
The safety net Washington doesn't trust was built in Beijing
Here's the part that should needle every US policymaker arguing for tighter open-weight controls. Andrew Ng told the Agentic AI Summit in Berkeley that he now finds open-weight models safer to work with than closed ones, not on principle but because OpenAI's and Anthropic's models refused to help him run a security review on his own tool, OpenWorker. He used Moonshot's Kimi K3 and Zhipu's GLM-5.2 instead. Hugging Face did the same thing last month, reaching for GLM-5.2 to investigate the very breach caused by an OpenAI model, because the US models it would normally trust couldn't tell a defender from an attacker.
Beijing, meanwhile, is tightening its own grip selectively. China revised its chip design protections this week. Its Commerce Ministry has been discussing curbing overseas access to frontier models with Alibaba, ByteDance and Z.ai, and Alibaba plans to start charging heavy users of its next open-weight model, according to Reuters — the free-and-open era of Chinese releases may already be ending, just as Washington debates banning the category outright. On a different axis, China has moved fast and hard: new nationwide rules on emotionally interactive AI forced ByteDance to shut Doubao's AI companion feature in July, with Alibaba and Tencent pulling similar features days later, ending relationships some users had maintained for over a year. Beijing will regulate ruthlessly when it decides a harm is real. It just hasn't decided cyber capability in its own open models is one yet, even as Dario Amodei spends weeks arguing exactly that. This week, American security teams were using those models to defend American companies, because the American alternative refused.
Two regulatory philosophies, tested in real time
The White House's near-final AI framework, discussed with executives this week per the Wall Street Journal, requires only voluntary government review for closed, proprietary US models hitting state-of-the-art cyber benchmarks. Open-weight US models are exempt entirely. Meta wasted no time exploiting that gap: the same day, Mark Zuckerberg published a 6,500-word case for open, lightly regulated AI, alongside a new open-weight model, Muse Glimmer, distilled to 30 billion parameters under an Apache 2.0 licence. Notice what stayed closed, though — Muse Spark, the more powerful model Glimmer was distilled from. Open the smaller thing, keep the frontier thing under control. That's the two-tier world Washington has now written into policy.
Brussels took the opposite bet. The EU AI Act's enforcement machinery went live on 2 August, and the European Commission has since published complaint and whistleblower tools letting insiders anonymously flag violations, backed by fines up to €15 million or 3% of global turnover. Seoul is hedging a third way. South Korea's government-backed race to build a sovereign foundation model is down to four contenders — LG AI Research, SK Telecom, Upstage and Motif Technologies — with a 200-citizen panel testing all four between 8 and 11 August before one gets cut this week. It's not Silicon Valley's terms or Beijing's. It's a mid-sized economy refusing to depend on either.
Capital doesn't wait for the argument to resolve
None of this slowed the money. South Korea and Taiwan each surpassed Japan in total exports for the first half of 2026 for the first time ever, Nikkei reported, both riding the chip boom. Leopold Aschenbrenner's fund, Situational Awareness, sold most of its public portfolio to Citadel as AI infrastructure stocks slid and its assets fell from $20 billion to $10 billion — yet still found $400 million for a chip manufacturing startup, Source Foundry. OpenAI completed a $7 billion tender offer buying back employee shares at an $852 billion valuation, having filed confidentially for an IPO in June; a tender like that usually means the public listing isn't imminent. Amazon is building a gas-fired plant in Pecos County, Texas, permitted to emit up to 33 million tonnes of CO2 a year for a single data centre, more than any operating US coal plant — the same week OpenAI wrote to Texas Governor Greg Abbott pledging "responsible AI infrastructure." Moody's, meanwhile, warned that banks racing into AI are building dependency on a small cluster of Silicon Valley providers, exposing the sector to cascading outages and price shocks as loss-making labs come under pressure to turn a profit; over 75% of City of London firms already use AI, and Lloyds has committed £13 billion to an AI strategy that includes £2 billion of cost cuts. Google used the week to quietly de-risk its own model-building, moving Demis Hassabis from day-to-day work to Chair of Google DeepMind and Alphabet's Chief Scientist.
What I'm watching next week
I want to see whether OpenAI's Astra pause holds once outside evaluators, including UK AISI, finish testing it, or whether commercial pressure reopens it the way Daybreak Red already has. I'm watching Alibaba's pricing move for signs the rest of China's open-weight fleet follows, which would remove the fallback US defenders are quietly relying on. And I'm watching Seoul's elimination result for what a government-run alternative to both Silicon Valley and Beijing actually looks like once it survives contact with users.
If you want this every morning in five minutes, the AI Storm Daily briefing is on the Hexalink blog, Spotify (https://open.spotify.com/show/033LojZEJj9VNX8b3Dm6VO) and Apple Podcasts (https://podcasts.apple.com/gb/podcast/ai-storm-daily/id6788420238).
Sources
- Responding to the next frontier of critical cyber capabilities — OpenAI — https://openai.com/index/responding-next-frontier-critical-cyber-capabilities
- Expanding Daybreak as the Cyber Defense Window Narrows — OpenAI News — https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows
- Putting frontier cyber models in more trusted hands — OpenAI News — https://openai.com/index/putting-frontier-cyber-models-in-more-trusted-hands
- As AI-led attacks multiply, OpenAI launches a new cyber model — TechCrunch — https://techcrunch.com/2026/08/10/as-ai-led-attacks-multiply-openai-launches-a-new-cyber-model/
- Tech industry is buzzing after a Claude agent hacked