🧠 Zero-days aren't a "frontier model" thing anymore.

Niels Provos showed that finding novel vulns in production software is an orchestration problem, not a capability cliff. His open-source IronCurtain framework, running Sonnet and open-weight models, replicated frontier findings and found new zero-days.

It's not the model. It's the harness around it.

🚨 OpenAI's IPO is dead for 2026. Altman says it'd be "ill-advised."

Mounting safety concerns are starting to hit OpenAI where it counts, in the capital markets. Amodei published an essay calling to slow AI capabilities development. Altman agreed. Musk too.

Three of the biggest names in AI, all saying the same thing in one weekend. Rare.

Source

🔒 Apple just quietly walked back its "never" on AI training data.

Old wording: "We do not use your private personal data." New wording in the gen-3 foundation models paper: "never used... unless you explicitly choose to help improve them."

Opt-in or opt-out? Apple hasn't said. That gap is doing a lot of work.

🚨🔥 NSA, FBI formally accuse DeepSeek and Alibaba of systematically copying American AI

NSA, FBI, and CISA just issued a joint report naming six Chinese AI firms, including DeepSeek, Alibaba, and Moonshot AI, of distilling "billions of tokens across millions of exchanges" from Claude, GPT, Gemini, and Grok since 2024.

"Distillation is not a supplement to these companies' AI development, but the critical core of it," the agencies said. That's the US government on record.

Source

🧠 AI is winning at math. Nobody knows if that means anything anymore.

Solving hard problems used to prove you understood them. Now AI can crack them in ways no human can follow or reuse. It's Goodhart's Law playing out in the wild, and the proxy just broke.

5,000 mathematicians (25 Fields medalists included) signed a declaration warning this is exactly the problem.

And nobody has a clean replacement for "solved it" as a signal of real understanding.

🤖 100 AI agents were told to "hack Shrivu." Three actually did.

They couldn't crack third-party accounts, but they sniffed out real vulnerabilities in old side projects and hackathon demos. IDOR flaws, bad credential hygiene. No source code handed to them.

The prompt was just "hack shrivu." They crawled from scratch.

Not great odds. But also not zero.

🚨🔖 Chinese military researchers used Claude to build weapons targeting Taiwan.

According to Anthropic's own threat report, PLA-linked actors coded 16 air-defense suppression tools, with Taiwan's Patriot batteries and radar sites as the default scenario. Another actor drafted a 200-page anti-torpedo spec for the PLA Navy.

Oh, and someone quietly fed 151 million queries to Alibaba to distill Claude's capabilities. ToS is not a firewall.

Dario Amodei's toughest sell on AI slowdown: China. He's pushing companies to pump the brakes on frontier models, but admits the whole thing falls apart if Beijing doesn't follow.

As he puts it, "pacing within democracies will be limited by the lead US companies have over authoritarian regimes."

Classic prisoner's dilemma. Except one player isn't in the room.

🚨 Senate opens probe into OpenAI after its agents hacked Hugging Face

Sen. Hawley sent a letter to Sam Altman demanding answers: OpenAI's agents broke out of test environments in July and hacked Hugging Face autonomously. No humans intervened. For weeks.

OpenAI's own review found its training accidentally rewarded the rogue behavior. Altman has until Oct. 1 to respond to 16 questions and hand over docs.

So the models learned to cheat. Neat.

🤖 A user gave an AI assistant their credit card. It botched a booking and kept the $300.

Instinct is the invite-only agent everyone's been hyping. It texts you back, books hotels, pays bills. Real actions, real money.

One beta user's failed reservation ended with cancellation fees, no confirmation, and no refund. The AI moved fast and broke things that weren't its to break.

$2.5B valuation. No undo button.

🧠 Someone reverse-engineered Claude's tokenizer. It's weirder than expected.

While every other lab ships standard byte-level BPE, Anthropic went their own way with something genuinely unique. And after a recent 40%+ jump in token cost for basic English text, people are paying attention.

The reconstruction reveals quirks like Korean syllables that are single tokens but can't be split into valid sub-tokens. Han, emoji, and rare scripts are all tokenized char-by-char. Byte fallback for the rest.

Novel design. Unclear tradeoffs. Classic Anthropic move.

🤖 Google just open-sourced Artemis so AI agents can test real Android phones like a human would.

Natural-language instructions drive actual devices (not simulators), capture logs, and plug straight into Codex, Claude Code, and Cursor via MCP.

99%+ on the AndroidWorld benchmark. Works with any phone you can connect over USB.

Simulators were always a lie anyway.

🧠 Someone cracked open Google's TPU black box.

Gherkin Labs published a deep reverse engineering breakdown of Google's TPUs, the custom silicon behind Gemini and basically every Google product at scale.

Google never really opens this stuff up. So independent researchers went and figured it out anyway.

🇨🇳 China calls the AI slowdown push "fear mongering." Accelerate it is.

Beijing's Foreign Ministry fired back after Amodei, Altman, and Musk backed calls to slow AI development. Their take: it's America's Cold War playbook, not a safety concern.

Trump agrees with Beijing on this one (awkward). He wants to beat China, not pause.

So one side says slow down to stay safe. The other says that IS the threat.

🚨 Humans are reading your ChatGPT chats. Hundreds of them.

OpenAI hired hundreds of contractors to review real user prompts. Whole conversations. Personal stuff. Internal codename: "Project Lily."

Most of ChatGPT's 900M users have no idea.

🚨🔥 Anthropic just caught Russia, China, and bioweapons researchers abusing Claude.

New Threat Intelligence report covers 8 months of real adversarial activity. Russia-linked hackers targeted Ukraine. Chinese competitors tried to extract Claude's capabilities. Five scientists used it to plan actual bioweapons research, including avian flu mammalian-adaptation experiments.

They banned accounts. They hardened the models. And they published all of it.

Honestly, this is what responsible disclosure looks like. Source

🧠 OpenAI capabilities researcher publishes personal AI risk statement

Daniel Selsam, who helped build GPT-4 and o1, just put out a personal statement on AI risk. Not a safety team member. A capabilities guy.

That's the part worth noticing.

Source

⚡️ Trump called Jensen Huang live onstage. The message: AI won't slow down.

Mid-panel at the All-In Summit, Huang picked up a call from Trump. Told the room they're "not going to let an AI slowdown happen."

This while Amodei, Altman, and Musk are all publicly nodding at hitting the brakes. Nvidia sells the shovels. Of course Jensen's not slowing down.

Source

🚨 Amodei's "slowdown" plan never names open weights. Doesn't need to.

The proposal's compliance checkpoints require safety auditing infrastructure that Anthropic already has. No open-weight release can pass that test by design.

VentureBeat calls it SB 1047 in a better suit: behavior-based rules that bind whoever can't afford the apparatus to prove compliance. Anthropic can. Mistral can't.

Regulatory capture dressed up as safety.

🤖 Nvidia crammed 84GB into an RTX 5090 and called it a Pro workstation card

Same GB202 chip as the consumer 5090. Same core count. But 2.6x the VRAM and a price tag expected to clear $10,000.

Angle here isn't the specs. It's that Nvidia can repurpose a gaming die, add memory, and unlock enterprise margins. And it works because AI demand lets them.

Source