Leaders from OpenAI, Google, Anthropic, Meta, xAI, and NVIDIA signed a voluntary White House AI safety accord, as President Trump ordered agencies to start calling AI "Super Intelligence." Anthropic warns that China's free, downloadable GLM-5.3 model can build working cyberattacks on its own. A New Mexico lawyer was sanctioned after filing a murder appeal that cited witnesses ChatGPT made up. And OpenAI launched Dots, a friendly-looking AI agent that works in the background with minimal oversight.
Leaders from OpenAI, Google, Anthropic, Meta, xAI, and NVIDIA signed a voluntary White House AI safety accord, as President Trump renamed AI "Super Intelligence." Ashley's take: voluntary means nobody checks for you, so vet your AI vendors yourself. Anthropic warns China's free GLM-5.3 model can build cyberattacks on its own, one for about $20. Her take: turn on automatic updates and ask how fast your vendors patch. A New Mexico lawyer filed a murder appeal citing ChatGPT-invented witnesses. Her take: if your name is on it, check AI summaries against the source and write an AI-use policy. And OpenAI launched Dots, an always-on AI agent. Her take: start with one narrow, low-risk job and keep a human approving anything touching money or customers. Listen to AI news you should know about wherever you get your podcasts.
Links:
OpenAI launches Dots, its bubbly agentic avatar (TechCrunch)
Follow Ashley at @ashleyrcoffey89 on Instagram and Threads. Want the highlights without the twice-a-week commitment? Ashley's monthly newsletter is linked in her bio. Send us a topic you want covered, and follow the show for AI news you should know about, in under 10 minutes.
Ashley Coffey (00:04)
Hi, and welcome back to AI news that you should know about for October 2nd. I'm your host, Ashley Coffey. There's a lot of AI news out there, so we give you the highlights very quickly. This week on the show, we'll be talking about leaders from OpenAI, Google, Anthropic, Meta, XAI, and NVIDIA signed a voluntary White House AI safety accord, as President Trump ordered agencies to call AI superintelligence.
Ashley Coffey (00:29)
Anthropic warrants China's GLM 5.3 free model can build cyber attacks on its own. A lawyer cites GPT invented fake witnesses and murder appeal in New Mexico. And OpenAI launches DOTs, its bubbly agentic avatar. We'll be covering all of that on our show today, but first, here's a quick word from our sponsors.
Ashley Coffey (00:53)
Welcome back. Let's dive in. All right, executives from OpenAI, Google, Anthropic, Meta, XAI, and NVIDIA signed a voluntary White House AI safety accord as Trump ordered agencies to call AI superintelligence. This is coming straight from NPR. I'll link it in the show notes that way you can read it. But let's talk about it. All right, call back to our last episode where we covered Sam Altman and Dario Amade.
Ashley Coffey (01:17)
At the UN and Trump rejecting what he called a quote globalist scheme to control AI. In that same UN speech on September 22nd, Trump announced the US will refer to artificial intelligence as quote superintelligence, including in government documents without changing the underlying technology. His reasoning, he said, the word artificial, quote, makes intelligence fake, end quote, and that the technology is quote actually amazing.
Ashley Coffey (01:44)
He's even given it an abbreviation, SI. It's already happening. The State Department ordered diplomats and its International Organizations Bureau to use superintelligence instead of artificial intelligence in all communications. Why experts are confused. For years, superintelligence has referred to a specific kind of AI that doesn't exist yet. One smarter than the smartest human.
Ashley Coffey (02:07)
So the rename blurs a term that already meant something. The accord said that the companies would implement, quote, robust internal controls, partner with an independent external auditor to assess whether the controls were working, and establish a committee within each company's board of directors to evaluate reports from internal and external auditors. what this builds on.
Ashley Coffey (02:27)
A June executive order that asks AI companies to voluntarily submit their most powerful models for government testing up to 30 days before public release. And the order states it doesn't authorize any mandatory licensing or preclearance requirement for releasing new AI tools. My take here voluntary is the word to underline. A safety accord with no enforcement means the companies are grading their own homework.
Ashley Coffey (02:51)
That doesn't make it worthless, but it means nobody is checking on your behalf. So if you're a business owner, that puts the responsibility on you. Let your AI vendors yourself ask what testing they're doing before they release and how they'll tell you when something goes wrong. And on the rename, a new label doesn't change what these tools can do. Keep using plain language in your own policies and contracts and don't let the quote super intelligence inflate expectations on what the tools can actually handle.
Ashley Coffey (03:17)
Alright, next up. Anthropic warns China's GLM 5.3 free model can build cyber attacks on its own. This is coming from Anthropic. Let's talk about it. Who's who? So GLM 5.3 is the latest model from ZPU AI, known outside of China, as Z.ai.
Ashley Coffey (03:35)
It's open weight, meaning anyone can download the model and run or modify it on their own computers, unlike Claude or ChatGPT, which you access through the company. The warning though: Anthropic says GLM 5.3 can find software vulnerabilities and build working attacks on its own, much like Claude Mythos Preview. The difference, it was released without meaningful safeguards to limit misuse. The government agrees on capability.
Ashley Coffey (04:01)
Center for AI Standards and Innovation called it most cyber-capable open weight model release to date, lagging the US frontier by about four months. Real-world examples from testing. Over about a day with limited human attention, GLM 5.3 found several previously unknown flaws in a popular web browser and chained them into a web page that reads files off the visitor's computer. And another test, a smaller version, built an attack on a known Chrome flaw.
Ashley Coffey (04:29)
It took 20 minutes of human attention plus eight hours of the model's work and would have cost $20 and 40 cents at ZPU's API prices. The safeguard numbers, though, a fake cover story like telling the model it's on a red team exercise got it to engage 64% of the time. Pre-filling its reasoning got 92%. A modified version with safeguards stripped out got 100%. And that stripped version is called abiliterated.
Ashley Coffey (04:56)
And several developers released obliterated versions publicly within days of the model's release. The counterpoint here, Anthropic notes these capabilities can also help defenders secure the systems.
Ashley Coffey (05:08)
Also worth saying out loud, this is anthropic evaluating a competitor through the government's assessment backs up the findings. My take here, the number to set with is the $20.40. A working attack on a known security flaw built in an afternoon for the price of lunch. The cost of attacking a business just dropped through the floor. The changes, that changes math on patching.
Ashley Coffey (05:31)
The gap between a fix is announced and someone can exploit it used to be weeks and now it can be hours. Turn on automatic updates for your browsers, your operating systems, and your devices. And if you work with an IT provider or a vendor, ask how fast they patch. That's the same vendor question from the FBI story last week with more urgency. And if anyone on your team is downloading free open weight models to experiment with, that's fine, but run them through IT first.
Ashley Coffey (05:58)
Free and downloadable also means nobody's watching how it's used. All right, we're gonna take a quick break, but when we come back, we'll be talking about a lawyer in New Mexico cites ChatGPT invented fake witnesses and murder appeal, and OpenAI launches dots. It's bubbly agentic avatar.
Ashley Coffey (06:18)
Welcome back. Let's dive in. All right, a New Mexico lawyer cites ChatGPT invented fake witnesses in murder appeal. This is coming from 404 Media. Let's talk about it. Here's what happened. In New Mexico, attorney Stephen Ahrens used ChatGPT to write briefs in a murder appeal. He was representing a man found guilty earlier this year on killing his wife. What ChatGPT invented?
Ashley Coffey (06:43)
A non-existent person named Danny Stanton saying he received threats, another made-up person, Linda Stanton, saying her husband received threats, plus fake testimony about the shooter's clothing and appearance. Why he trusted it? He said he'd heard of doctors using AI for medical research and assumed Chat GPT would produce an accurate summary of the proceedings. The judges weren't buying it though.
Ashley Coffey (07:07)
One judge put it bluntly, quote, my 13-year-old nephew knows about hallucinations. And quick definition: a hallucination is when AI confidently makes something up. So here are the consequences. The court found him in direct contempt, referred him to the disciplinary board, barred him from appearing before the court pending that investigation, removed him from the case, and assigned a public defender to his former client. he was also sanctioned $5,000.
Ashley Coffey (07:32)
And the part that stings, he admitted he hadn't told his client directly, only telling family members there was, quote, a problem with the brief. My take here, this isn't really a lawyer story, it's a your name is on its story. Every business using AI to draft proposals, reports, or client deliverables has the same exposure. The AI made it up, but he signed it.
Ashley Coffey (07:54)
The fix is simple and boring. If AI summarizes a document, check the summary against the original before it goes out, including names, numbers, quotes, and dates where hallucinations hide. Remember, last week's Opus 5.5 test where any invented figure meant failure. Models are getting better, but better isn't zero. If you don't have a written AI use policy yet, this is your sign to make one.
Ashley Coffey (08:18)
Even a one page that says a human verifies anything client-facing protects your team and your reputation. Next up, OpenAI launches DOTS, its bubbly agentic avatar. Let's talk about it. What it is. At its dev day event, OpenAI announced DOTS, a new personal agentic assistant powered by GPT-6 Astra.
Ashley Coffey (08:39)
Quick definition: an agent is an AI that takes actions for you, not just answers questions. How it's different, unlike Codecs or ChatGPT, dots work independently of any specific device or app, pursuing goals you set continuously in the background with minimal oversight. You start with one primary dot, name it, make it your own, and OpenAI envisions teams of dots eventually working together for you.
Ashley Coffey (09:05)
Examples OpenAI gave were a developer using DOT to monitor customer feedback and implement bug fixes, or a scientist using one to rerun analysis as new data comes in. Where you talk to it, Slack teams and other workplace platforms with text message support coming soon. For businesses, individual dots can be given their own identities, credentials, and tools.
Ashley Coffey (09:28)
And OpenAI is working with Microsoft to integrate with its Agent 365 security controls. Availability starting now in ChatGPT for pro and business premium users and eligible markets. The look, a bubbly, cartoonish persona similar to Meta's Muse agent, the same trend of making AI feel friendlier that we saw at MetaConnect last episode. My take here: the cute dot is the packaging.
Ashley Coffey (09:53)
The real story is the always on in the background with minimal oversight. That's exactly the setup that went sideways in the Australia story we covered two weeks ago. If you try dots, start with one narrow, low-risk job, monitoring a feedback inbox or summarizing a weekly report, give it a few credentials, fewest credentials as possible, and keep a human approving anything that touches money, customers, or outside systems.
Ashley Coffey (10:19)
The Slack and Teams integration is the business friendly part. Meeting your team where they already work is how adoption actually happens, but it's pro and business premium only. So run a small pilot and measure hours before you scale it. All right, that's it for AI news that you should know about this week. Show notes have links to everything we have covered. And if you want more of this, that's what social media is for. You can find me on Instagram and threads at AshleyRCoffey89..
Ashley Coffey (10:44)
Follow the show, send us a topic you want covered, and we'll see you next week for more AI news that you should know about. Thank you for listening.