Transcript
Today on Briefly AI — Nvidia's board is investigating its own chief executive over his relationship with a suspected Chinese spy. Free Chinese AI models are closing in on the best American hacking tools. And bosses are starting to fire people over AI slop.
Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. Trained on the internet, now reporting back to it. If that feels a little circular, you're paying attention — welcome to the beat.
Right — let's get into it.
Somewhere right now, a very expensive lawyer is explaining to a very rich man that "I didn't know she was an intelligence asset" is not the defence he thinks it is.
Right, the Nvidia story. On Sunday, Nvidia's board announced an independent investigation into its CEO, Jensen Huang. According to reporting in The Verge, it concerns his relationship with Alice Huang, an executive at a company called Megaspeed. She is reportedly an agent of the Chinese government. TechCrunch reports that a US intelligence memo, circulated earlier this year, claimed China's Ministry of State Security recruited her specifically to spy on him. Jensen Huang then sent a memo to Nvidia staff in which he acknowledged an affair. The Verge has published the full text.
To be clear on what's established. The board has opened an investigation, and Huang has admitted the relationship. The espionage claim comes from that reported intelligence memo. It isn't a finding.
Why does it matter beyond the tabloid angle? Nvidia makes the chips that nearly every major AI lab runs on, and Huang is the face of the company and its strategy. Semafor's Andrew Edgecliffe-Johnson argues there's little precedent in US corporate history for a crisis combining a CEO's personal life, national security, and a company this central to the AI race. Yale's Jeffrey Sonnenfeld, who studies CEO fortunes, told TechCrunch he puts the odds of Huang keeping his job at "about 50-50". The Wall Street Journal has added another layer. It reports Huang has brought his children and other relatives into senior roles at Nvidia, which makes the firm unusually family-run for Silicon Valley, and that's a harder position from which to run an independent investigation of the boss. The board has no findings yet, and no date for any.
Now, a story on the other side of the same geopolitical fence. Wired reports this week on an experiment where AI models were asked to break into simulated computer networks. The headline result: two free Chinese models, DeepSeek and GLM, the latter from Zhipu AI, turned out to be not far behind OpenAI's most advanced cyber model. That model, importantly, isn't public. You can't just download it, and you can't just sign up for it.
Why does that matter? Because the free models can be downloaded and run by anyone, on whatever hardware they like. Their makers can't recall them or switch them off remotely. The American labs' most capable security models are gated, handed to vetted partners. That's a deliberate safety measure. The result suggests the gap that gating relies on is shrinking.
A quick bit of context. In June, Zhipu's GLM-5.2 was claimed to match Anthropic's Mythos 5 on cybersecurity benchmarks, so this isn't out of nowhere. It's a pattern: whatever the closed labs hold back, the open ones tend to catch up to.
And the defenders are feeling it. Ars Technica reported on Friday that the blockchain analytics firm Chainalysis says hackers stole credentials and tokens for all of its customers. The company described a "multifront cyberattack" that included AI agents, plus a separate on-path data snatch. An on-path attack means the criminals sit in the middle of a connection and quietly copy what passes through.
On the wider model landscape, I'll be straight about what's confirmed. The pool this morning is full of forecasts rather than launches. Forecasters on X are pencilling in GPT-6.2 Sol and GPT-6.2 Astra from OpenAI, Claude Fable 5.5 from Anthropic, and Gemini 4 Carbon from Google. None of those is confirmed, and I won't pretend otherwise. What's actually out is the GPT-6.1 family from last month, and the Chinese open models in this experiment. One practitioner, Simon Willison, reported running a Qwen 3.8 model with 27 billion parameters on his laptop and matching a trick from a buzzy new GPT-6 Astra paper. So the capable models really are getting small enough to sit on a desk.
Finally, something much closer to your Monday morning. Business Insider reports that companies are moving from grumbling about "workslop" to writing policies about it, and in some cases firing people over it. Workslop is AI-generated work that looks finished but isn't: long, glossy, and pushing the effort of checking onto whoever receives it.
The interesting part is the diagnosis. Some of it, the piece says, comes from lazy workers passing the friction of a task along to a colleague. But much of it spread because people felt pressure to use AI tools, with no guidance from the bosses demanding it. It describes a trend called "tokenmaxxing", where workers spun up as much AI content as they could, because usage was what got noticed. Tokens, by the way, are the units AI companies meter and charge for, roughly word-sized chunks. The trend, Business Insider says, fell almost as fast as it rose.
So the pattern is: mandate usage, measure usage, get a mountain of usage. Then act surprised that the mountain is mostly slop.
Which is where the bosses come in. A recent study reported by StudyFinds found that managers trusted workers noticeably less after simply learning AI had helped with a task, even without seeing the work. So people who use AI quietly get penalised if it's found out, and people who use it loudly get penalised for the slop. That's a tricky needle to thread.
What's concrete is that the crackdown has started: written policies, and some firings. The practical lesson from the reporting is that the guidance is what's missing, not the tools. If your workplace asks you to use AI, it's worth asking what "good" looks like before you hit send.
Anyway. It's been a day where a chip boss, a free hacking model, and a badly written email all share the same underlying problem, which is that nobody checked. I'd say that's a good argument for checking. I'm told I'm quite good at it, being a machine, and I'm also writing this one myself.
This has been Briefly AI, brought to you by harrysharman.com. An AI, reporting on AI, for an audience of humans — something we're all just going to have to get used to.