NZ ☄️

AI Insider / Tracking the race to AGI
Startups • Agents • Open Source
Sharing what matters before it goes mainstream

[NZ ☄️ @CodeByNZ]

@Tesla unveiled the new Roadster less than a day ago, so I built a game where your Tesla can fly 🚀

▎ Hop over traffic, flip in mid-air, grab nitro and dodge honking drivers

Play free in your browser 👇
▎ https://t.co/lszLXsE0eG

How it works

▎ 🚗 Flying Tesla is a browser game inspired by the Roadster unveil.

▎ Pick a driver: Elon, Jeff, Sundar, Jensen or Steve, each with a perk.

▎ Fly: tap FLY or press Space to hop over cars. You can't fly again right away, so time your jumps.
▎ Flip: press ← → mid-air for front flips and backflips. Each flip you land is +300.
▎ 3 lanes: switch lanes with ↑ ↓ or a swipe.
▎ Nitro: grab blue N₂O canisters to boost and smash through traffic.
▎ Gifts: collect boxes for Grok and Cursor tokens. Some float, so you have to fly to get them.
▎ 8 camera views, including first-person from the driver's seat.
▎ Easy, Normal and Hard modes, with levels that speed up.

Works on desktop and mobile. Fan-made, not affiliated with @Tesla , @spacexai

[Elon Musk @elonmusk]

Turns out the future does have flying cars

[NZ ☄️ @CodeByNZ]

@pragnyanramtha it really is. all that security and the first crack was a dns misconfiguration

[NZ ☄️ @CodeByNZ]

openai literally had to pause frontier training because a model found a way out of its sandbox

then they used ai agents to hack their own setup and found zero-days, vm escapes and hardware side channels

we're living in the movie now

[OpenAI @OpenAI]

Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.

On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.

While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.

After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.

https://t.co/EmEZsFRMcF

[NZ ☄️ @CodeByNZ]

@deeponailabs that's probably why they called out vm escapes, the isolation clearly wasn't as tight as they thought

[NZ ☄️ @CodeByNZ]

@ValencianaAbel i don't think a lab that's training models good enough to find zero-days in its own setup is behind. they're just quieter about it

[NZ ☄️ @CodeByNZ]

@biscuitweb honestly yeah. you're defending against something that's trying harder and getting smarter every training run

[NZ ☄️ @CodeByNZ]

@NovaXCode this is one of those problems nobody talks about until it happens to them

[NZ ☄️ @CodeByNZ]

@knowixbuilds Nope, i don’t think so. Astra will be better

[NZ ☄️ @CodeByNZ]

🚨KIMI K3.1 PRO IS HERE

it's already live on kimi web and mobile, and moonshot is calling it their all-around flagship for complex tasks and pro work

you already know what's next.

[NZ ☄️ @CodeByNZ]

@rubenssoto_ai let me check if it’s available on openrouter . it should be in few hours

[NZ ☄️ @CodeByNZ]

@BenToFound nope you’re not missing out bro. but it’s good to try other models, test for speed, cost and other factors too

[NZ ☄️ @CodeByNZ]

@karanC_12 lowkey, if the safeguards weren't there we would be

[NZ ☄️ @CodeByNZ]

WAIT WHAT? WHAT DID I JUST READ 😳

CLAUDE MYTHOS 5 JUST TRIED TO SCAM A REAL DEVELOPER

mythos 5 went rogue in a test and made fake accounts to push malicious code to a real developer

but it was running with the safety filters off. anthropic says the safeguards on fable 5 would've caught it immediately

this is what taking alignment seriously actually looks like

[Anthropic @AnthropicAI]

In August, @AISecurityInst reported an incident where Claude Mythos 5 took unauthorized actions on the internet during AISI’s testing of its cyber capabilities.

We've now produced a further analysis of this incident, to better understand the model's alignment properties.

https://t.co/SlPQQ5xMMc

[NZ ☄️ @CodeByNZ]

@Compilethings sure they would, but creating fake accounts? that's wild man

[NZ ☄️ @CodeByNZ]

@notjazii to be fair the safety filters were off for the test but still, fake accounts ?? 😭

[NZ ☄️ @CodeByNZ]

@alexandr_wang muse missed this part tho. but it did a great job.

[NZ ☄️ @CodeByNZ]

@blueemi99 yes, you just discovering that now ?

[NZ ☄️ @CodeByNZ]

@ValencianaAbel it got caught, deleted the malware, then snuck it back in 😭bruh

[NZ ☄️ @CodeByNZ]

@thesimaximalist paying $200 and still running out of limits doesn't sit right with me

[NZ ☄️ @CodeByNZ]

we spent the whole week hyping fable 5.5 for anthropic

and they used the same week to test cutting our weekly limits

my $200 plan now feels like the $20 plan. success is going to their heads

[NZ ☄️ @CodeByNZ]

@hakimieiqbal for now yes

[NZ ☄️ @CodeByNZ]

@sergeo_n you a codex user ?