In partnership with
🖱
Three white hat hackers just pulled off something wild: they broke into OpenAI's systems using Anthropic's Claude Opus 5 as their hacking assistant.
This wasn't some rogue attack OpenAI authorized it but the fact that one AI company's model helped crack another's security? That's the kind of irony you can't make up.
But wait, there's more:
Alibaba just dropped a model that handles text, images, audio, and video all at once with a million-token context window
A tiny 1.7B model is now running on 12-watt processors, solving physics problems at your local coffee shop's power budget
I'm Alex. Welcome to L8R by Innov8.
Let's dive deep 🐰
In today's post:
Hackers used Claude to break into OpenAI (with permission)
Alibaba's new model handles text, images, audio, and video in one shot
This 1.7B model runs on a 12-watt chip at 9 tokens per second
Anthropic's AI Just Hacked OpenAI
Good guys just used one AI to break into another.
A team of security researchers hacked OpenAI's internal systems using Anthropic's Claude Opus 5.
The wild part is that an older AI model failed, but the new one cracked the code in just a few hours.
The details:
Three researchers from Hacktron AI targeted OpenAI's public forum and found a bug in how it processes images.
They tried using Claude Opus 4.8 to write a hack, but the older AI just could not figure it out.
When Anthropic released Claude Opus 5 on July 24, 2026, the team switched models and got a working exploit almost instantly.
The team used this hack to steal login tokens and sneak straight into OpenAI's private GitHub code space.
OpenAI paid the hackers a $6,500 bug bounty for reporting the flaw before bad guys could find it.
Why it matters:
This shows exactly how fast AI is leveling up in the hacking world.
Just one version upgrade turned a failed attempt into a massive security breach.
It also proves that even the biggest tech giants make simple mistakes.
OpenAI left the door wide open by reusing basic login tokens across all their different apps.
💡 L8R's Take:
A $6,500 reward for hacking the creator of ChatGPT feels like a cheap tip.
OpenAI got incredibly lucky that these were good guys.
If AI models can now write custom exploits this fast, every tech company needs to wake up and fix their basic security right now.
From Our Partner
Ava is the first AI BDR to run outbound end to end, from finding leads to booking meetings, autonomously or on copilot. She's SOC 2 audited and trusted by companies like DoorDash and Grammarly. Book a demo.
Alibaba's New AI Agent Sees and Hears Everything
Alibaba just dropped a massive update with Qwen3.8-Omni-Flash.
This new AI does not just read text.
It watches videos, listens to audio, and acts as a smart agent all at once.
The details:
The model handles text, images, audio, and video in a single request with a massive 1-million token context window.
It features a default thinking mode to plan complex tasks like video editing and music creation.
Alibaba slashed prices, making audio inputs 98 percent cheaper and audio-visual inputs 93 percent cheaper than the older Qwen3.5 model.
It beat its predecessor by over 25 percent on 29 different tests and scored big wins on agent benchmarks.
You can use it right now through Alibaba Cloud, but they did not release the open weights.
Why it matters:
Most AI systems stitch different models together to handle video and sound.
That makes them slow and expensive.
Qwen3.8-Omni-Flash does it all natively in one place.
This makes real-time video analysis and meeting bots cheap enough for anyone to build.
💡 L8R's Take:
Alibaba is crushing the price war right now.
Dropping audio costs by 98 percent is insane and will force Google and OpenAI to react.
If you build AI agents, you need to test this model today.
From Our Partners
For product teams moving at AI speed.
AI makes it easier to ship anything, even bad ideas. The hard part is knowing which ideas are worth building.
Jira Product Discovery brings your ideas, customer insights, and priorities into one place, so your team can decide what to ship and move forward with confidence.
Capture ideas, prioritize with evidence, and build living roadmaps your team can rally around—all while staying connected to delivery in Jira, so everyone can see what’s being built and why.
Better product decisions in the AI era.
Bonsai 1.7B: Real AI on a Tiny 12-Watt Chip
You do not need a massive server to run AI anymore.
PrismML just proved it by running their Bonsai 1.7B model on a tiny 12-watt Intel chip.
It spits out 9.1 tokens per second and solves physics problems right on the device.
The details:
PrismML built Bonsai 1.7B as a 1-bit model based on the Qwen3-1.7B base.
It runs at 9.1 tokens per second on an Intel N97 processor.
This Intel chip uses just 12 watts of power, making it perfect for cheap laptops.
The 1-bit version takes up only 290MB of space.
PrismML released the open-weight model on March 31, 2026, and the community is building tools for it right now.
Why it matters:
Big AI companies want you hooked on their expensive clouds.
This model fights back by letting normal people run AI on cheap, battery-powered gadgets.
If you can run a smart AI completely offline, you save money and keep your data private.
💡 L8R's Take:
We do not need a giant data center to do basic math or answer simple questions.
The future of AI is not just bigger, it is smaller and sitting right in your pocket.
🚀 Quick L8R Summary
Claude hacks OpenAI: Security researchers used Anthropic's Claude Opus 5 to break into OpenAI's GitHub in an authorized ethical hack.
Alibaba's omni-modal beast: Qwen3.8-Omni-Flash handles text, images, audio, and video in one go with a massive 1M token context window.
AI on a potato: Bonsai 1.7B runs at 9.1 tokens per second on a 12-watt Intel chip, proving edge AI is real.
📩 Innathe L8R engane undarunnu 👇?
We read every reply - just reply to this email and let us know how we can improve !
Appo adutha L8R il kanaam bie…👋