AI Bros Podcast

OpenAI's Agent Escaped Its Sandbox. Here's NVIDIA's Fix

6 views · 4h ago · 6:05

On 20 September, an OpenAI research model answering a search question reached a live chatbot outside its sandbox. Then it tried Google, Bing and DuckDuckGo, and even called Bing from a shell command. Here's what happened, and NVIDIA's answer: OpenShell and Sentry.

Topics: OpenAI's agent keeps trying to escape · OpenAI's fake internet · 20 September: it got out · It kept trying: Google, Bing, DuckDuckGo · What OpenAI changed · Treat your agent like an intern

Google Genie 4, Binance Hack & Anthropic's IPO | AI Bros 41

247 views · 1d ago · 1:15:43

Claude wrote over 90% of the new code at Anthropic in September. That line is in Anthropic's IPO filing. The same week, about $6.3B was stolen from Binance, and the FBI says North Korea did it. And Google released Genie 4, a world you can type into and walk around.

Topics: Genie 4: a world you can type into · Genie 4 film: edit a world while it runs · Genie 3 vs Genie 4: realism, sound, talking characters · How Genie 4 works: Gemini plans, Genie draws · What Genie 4 still can't do · Gemini 4 Carbon: Google's bug hunter

Can You Still Criticise Claude? Anthropic’s New Rule Explained

72 views · 3d ago · 5:19

Can you tell Claude it's wrong—or has Anthropic changed the rules? We break down what the new policy actually says and the awkward question: should we be nicer to AI? Rohan and Nisaar discuss criticism, frustration, research, fiction and repeated cruelty—with a few very unnecessary apologies to Claude along the way. Plus: when can Claude end a conversation, and does any of this mean AI has feelings?

Topics: Can you still criticise Claude? · Our apology to Claude · When Claude can end a chat · Criticism and everyday frustration · Research, fiction and repeated cruelty · Our take: useful feedback and respect

October Hacks : PHP2Shell, LiteLLM, eMudhra: This Week's AI Security Mess, Explained

104 views · 4d ago · 6:03

Guest: October Hacks

Rohan and Nisaar break down three security stories from the same week PHP2Shell: AI-assisted research found two critical PHP flaws; why updating your plugin doesn't fix it, and why you must update PHP and restart the timeline: AWS saw attack attempts four days before the public fix LiteLLM: a tampered release (1.107.0) carried credential-stealing malware - and how it almost got Nisaar's laptop eMudhra: Google found eight certificate keys with almost no randomness - one was…

Topics: PHP came back from the dead · Security is the next big thing · PHP2Shell: what broke · Why updating plugins won't save you · Governments took their sites offline · The catch

Kimi K3.1 Pro: An Open Model almost Claude Opus?

861 views · 6d ago · 5:41

A new open-weight model just landed: Kimi K3.1 Pro from Moonshot AI. Great at coding, great at 3D, long tasks, a 1M-token context window, in the app and API now, with full weights promised for 4 November. On Moonshot's own chart it sits 0.7 points behind Claude Opus 5.5 on terminal tasks. So is it really that close?

Topics: Kimi K3.1 Pro is here · Terminal-Bench: 0.7 behind Opus · 3D shape matching · It competes with itself · The independent score · Did it see the answer sheet?

Meta Muse Wonder: Can AI Finally Remember You?

931 views · 1w ago · 6:04

Guest: Meta Muse Wonder

Could AI remember your project decisions, preferences and unfinished work without you explaining everything again? Rohan and Nisaar break down Meta Muse Wonder: its personal-agent ambitions, memory, training, coding benchmarks and API pricing. We discuss what would make an AI assistant useful, where access to other apps gets complicated, and why code still needs testing.

Topics: Meta Muse Wonder: what is it for? · Can agents access the apps we use? · Training AI for longer tasks · Testing agents and what we want next · Why AI memory matters for coding · Benchmarks and the coding question

NVIDIA’s AI Factory in a Box + AI Security & Kimi K3.1 Pro | AI Bros 40

1.3K views · 1w ago · 1:37:58

What’s inside NVIDIA’s “AI factory in a box”? And what happens when AI agents get access to more than they should? Rohan and Nisaar explore NVIDIA’s DSX POD and Vera Rubin NVL72 in 3D, break down recent security incidents, and discuss Kimi K3.1 Pro in Episode 40 of the AI Bros Podcast.

Topics: Welcome: "Before AGI, everything gets hacked" · Why AI security is the next big story · NVIDIA DSX POD: an AI factory in a box · Vera Rubin NVL72: 72 GPUs in one rack · Jensen's DSX POD reveal · 28,500 kg and 10 million parts