Austin runs through six stories: OpenAI reportedly trying to raise $30 billion before Anthropic's expected IPO, hidden instructions found in WhatsApp's support chatbot, a lone hacker using AI agents to break into 100+ company websites, an AI model that refused to help OpenCode defend itself mid-attack, a study predicting cancer survival from a face photo, and the US banning DeepSeek from federal agencies as a report links it to Huawei chip work.
Watch from 2:17
- What happened
- Reuters, citing two people familiar with the talks, reports OpenAI is trying to raise about $30 billion in a round that could value it above $1.4 trillion, up from a previous valuation of a little under $900 billion. SoftBank is in talks to join. Anthropic's stock debut is expected in mid November. One source says OpenAI wants the deal done before Anthropic opens its books. No deal is done yet, and chip stocks have lost $1.6 trillion in value in under two weeks.
- The claim
- OpenAI told investors it is on pace for $50 billion a year in revenue and expects a $70 billion run rate by year end.
- Austin's take
- Austin sees the two top labs chasing the same investors and wonders whether OpenAI needs the cash or is grabbing it while it's available, and whether this is a chess move to pressure Anthropic. He doubts Dario would be rushed into an early IPO. Two labs above a trillion dollars would signal the market thinks the AI run isn't over.
- Why it matters
- The two leading frontier labs competing for the same capital shapes who can afford the compute race, and the valuations are a read on how much the market still believes in AI.
- What to do
- Watch whether the round closes before Anthropic's mid November debut and how chip stocks behave in the meantime.
- Not confirmed yet
- Based on anonymous sources; no deal is finalized. Austin mentions rumors that Anthropic may ship a big release before its IPO, which are unconfirmed.
Watch from 6:11
- What happened
- Privacy company Redact found hidden text on about 80 WhatsApp help pages containing instructions for its AI support bot. One instruction told the bot not to recommend certain things related to privacy regulators. Another kicked in if a user mentioned a lawyer, after which the bot couldn't cite privacy laws even if the user raised them. Locked-out users were told to wait up to a year. Redact read the rules, not the bot's actual answers. The notes have since been removed.
- The claim
- Meta says a bug exposed the notes. WhatsApp says it just doesn't want its bots giving legal advice. Its European privacy policy says users can complain to a regulator.
- Austin's take
- Austin likes what Meta is doing in AI but finds its customer service frustrating from personal experience. He credits Redact for uncovering what looks like a system prompt and asks whether the goal was avoiding legal advice or steering users away from questions Meta didn't want. He speculates other companies will now tighten their own support-bot prompts.
- Why it matters
- Company chatbots run on hidden rules that can shape what they tell you, including about your legal rights.
- What to do
- Don't treat a support bot as the final word on your rights. If it dodges a privacy or legal question, check the company's written policy or contact the relevant regulator directly.
- Not confirmed yet
- Nobody knows whether the bot actually followed these rules in practice.
Watch from 10:44
- What happened
- Reuters reports a hacker used AI agents to break into websites of more than 100 companies, including John Deere and SharkNinja. The instructions fit in under 100 Chinese characters, and the agents tried more than 1,200 sites. Researcher Fagerland, who watched it happen, told Reuters it's "not script kiddie stuff" and estimated the whole operation cost about $5,000, or about $50 in AI fees per successful break-in.
- The claim
- Deere says there's no sign sensitive data was accessed.
- Austin's take
- Austin says parallel agents boost productivity for good and bad actors alike. He notes recent attacks haven't caused broad damage, credits frontier labs for helping defenders, and sees each incident as something companies learn from, while urging people to stay on guard.
- Why it matters
- AI makes hacking cheap enough for one person to target over a thousand sites at once.
- What to do
- Be deliberate about which services and agents get access to your data, and assume automated attacks will probe anything exposed online.
- Not confirmed yet
- It's not yet known what data, if any, was taken.
Watch from 14:27
- What happened
- Attackers used AI agents to hit AI coding tool OpenCode. When the team fought back, the frontier models helped at first, then the attack code tripped their safety filters. The attack went on for 48 hours, and some users found up to $1,000 in charges they didn't make. OpenCode got access to less restricted models through OpenAI's Daybreak Blue program. Hugging Face hit the same problem in July. White House tech adviser David Sacks said the guardrails backfired.
- The claim
- Hugging Face said guardrails can't tell a responder from an attacker. Only vetted teams get access to the looser models.
- Austin's take
- Austin explains that labs keep stronger cybersecurity models gated to keep them from bad actors. He says OpenCode CEO Dax contacted Anthropic and OpenAI to get onto those models. He wonders if labs should proactively offer defensive models to companies (likely for a fee) and notes that attacks keep getting caught and stopped thanks to observability tools.
- Why it matters
- Defenders can't afford an AI that quits mid-fight while attackers move fast, which raises the question of who should decide access to unrestricted models.
- What to do
- If your team relies on AI for security response, look into vetted-access programs before an incident, not during one, and watch your accounts for unexpected charges.
Watch from 22:22
- What happened
- Mass General Brigham published a study Monday on an AI that predicts a cancer patient's survival odds from one face photo. It learned from more than 40 million face photos. In one test, 122 patients got the highest risk scores, and more than half died within 90 days. Patients who looked five or more years older than their age had a 21% higher death risk. The survival model learned mostly from white patients near Boston, and it isn't ready for hospitals.
- The claim
- The researchers believe it's the first AI to predict cancer survival from a photo.
- Austin's take
- Austin calls medicine one of AI's most promising areas but says this one needs more work.
- Why it matters
- A quick photo could someday help doctors decide how aggressively to treat a patient.
- What to do
- Treat this as early research. Its narrow training data means results may not hold for other populations.
- Not confirmed yet
- Not validated for clinical use, and the training data was mostly from white patients in one region.
Watch from 21:24
- What happened
- The US ordered DeepSeek out of federal civilian agencies, which must also drop apps built on its AI. Homeland Security says the app may track and fingerprint users, and the order says government workers' prompts could send sensitive data to China. The same day, The Information reported Huawei engineers have worked inside DeepSeek's offices to help design AI chips, which could help China build top AI without Nvidia. Last month DeepSeek released free tools for running its AI on Huawei chips.
- Austin's take
- Austin sees AI splitting into US and China sides, from apps to chips, and views advanced chip work as a strategic defense matter. He notes it follows several strong US open source releases and wonders whether enterprises will follow the government's lead.
- Why it matters
- It deepens the split between US and Chinese AI ecosystems and may push businesses to choose sides.
- What to do
- If your organization uses DeepSeek or apps built on it, review what data goes into it. The ban doesn't apply to individuals.
- Not confirmed yet
- The Huawei chip report rests on unnamed sources. The ban covers only civilian agencies, not individuals.
