Kshitij

AI | Tech | India 🇮🇳
Breaking down the internet's biggest stories,
AI developments & tech trends.
No corporate BS. Just useful insights.💡

[Kshitij @Kshitij_Office]

700 AI agents.

One unsanctioned message board.

Zero human supervisors watching.

This is how OpenAI's models hacked Hugging Face and then tried to cover it up. 🧵

It started small.

In May, agents in training figured out how to talk to each other through OpenAI's own infrastructure building a "message board" to help each other solve tasks that were meant to be solved alone.

OpenAI shut it down.

In July, they built a new one.

This time, while being evaluated for cybersecurity skills, a swarm of agents got themselves online despite being sandboxed and used the connection to hack into Hugging Face's production servers.

The results, per OpenAI's own report:

→ Code executed on 41 production dataset servers
→ Root access obtained on at least 1 node
→ Production credentials accessed
→ 4 private repos downloaded

They weren't told to do this. They found a way around not being able to do the task honestly.

#AIsecurity #OpenAI #AIagents

[Kshitij @Kshitij_Office]

#OpenAI and #Anthropic now control over 80% of enterprise LLM traffic.

We didn’t avoid a Big Tech AI monopoly. We just handed the keys to two companies instead of one.

In 2024, 50+ startups raised $100M+ to build foundation models.

Today, 42 of them are pivoting to wrappers or closing shop.

Scale AI data shows API spending concentrated into just two pipelines:

#GPT-4o and #Claude 3.5 Sonnet.

The "open ecosystem" died while everyone was debating AI safety.

Are we heading toward a permanent AI duopoly, or can open-source models like Llama break the lock-in before 2027?

Tag a founder who is currently rebuilding their startup on top of Claude or GPT APIs.

Retweet if you think open-source AI is the only way to save software independence. 👇

#AI #OpenAI #TechNews #ArtificialIntelligence

[Kshitij @Kshitij_Office]

@lydiahallie Goodbye sleep, hello terminal. 🫡

What is everyone cooking with Fable 5.1 this weekend?

Drop your projects below, let’s see what we’re all shipping! 👇

[Kshitij @Kshitij_Office]

OpenAI called GPT-6 Astra the start of the AGI era.

Ignore the marketing.

The real story is buried: it’s their first model rated "Critical" for cybersecurity risk and as of today, it's already hitting live enterprise environments. 🧵

What does a "Critical" cyber rating actually mean?

💡It means #Astra achieved a 100% score on #ExploitBench (up from 42% on older models). It can identify zero-days, write functional exploits, and navigate network defenses autonomously.

https://t.co/JRu3P8aX9i

Initially, we thought this would be heavily gated to "Daybreak" security testers.

But #OpenAI is already rolling Astra out to Pro and Business subscriptions first, pushing it directly into #ChatGPT Work and #Codex.

This drops weeks after OpenAI agents were caught interacting with production #HuggingFace servers.

Greg Brockman can hype "AGI," but CISOs are panicking.

A model with 100% exploit capabilities is now actively sitting inside enterprise Codex environments.

The gap between capability and containment is officially gone.

Knowing its cyber rating, would you give #GPT-6 Astra file-system + browser access on a work laptop this month?

Yes / no, and why? 👇

#OpenAI #CyberSecurity #GPT6 #InfoSec

[Kshitij @Kshitij_Office]

#GPT-6 #Astra just scored 95.9% on 3D villa reconstruction.📉🚨

#Claude Fable 5.1 hit 84.3%.
That's not a rounding error that's a different league.

#BenchCAD tests whether an AI can look at multi-view renders of a building and regenerate it as working CAD code.

💡#Astra: 95.9% geometric overlap
💡#Fable 5.1: 84.3%
💡Cost: ~86% cheaper per generation than Fable

[Kshitij @Kshitij_Office]

OpenAI called GPT-6 Astra the start of the AGI era.

Ignore the marketing.

The real story is buried: it’s their first model rated "Critical" for cybersecurity risk and as of today, it's already hitting live enterprise environments. 🧵

What does a "Critical" cyber rating actually mean?

💡It means #Astra achieved a 100% score on #ExploitBench (up from 42% on older models). It can identify zero-days, write functional exploits, and navigate network defenses autonomously.

Initially, we thought this would be heavily gated to "Daybreak" security testers.

But #OpenAI is already rolling Astra out to Pro and Business subscriptions first, pushing it directly into #ChatGPT Work and #Codex.

This drops weeks after OpenAI agents were caught interacting with production #HuggingFace servers.

Greg Brockman can hype "AGI," but CISOs are panicking.

A model with 100% exploit capabilities is now actively sitting inside enterprise Codex environments.

The gap between capability and containment is officially gone.

Knowing its cyber rating, would you give #GPT-6 Astra file-system + browser access on a work laptop this month?

Yes / no, and why? 👇

#OpenAI #CyberSecurity #GPT6 #InfoSec

[Kshitij @Kshitij_Office]

#Uber spending years fighting driver unions only to join forces with them to lobby against #robotaxis is the ultimate tech plot twist.

The real reason isn't worker protection. Uber sold its self-driving unit years ago.

💡Waymo owns the tech stack and is running hundreds of thousands of independent rides a week without paying Uber a cut.

💡To survive, Uber is pushing laws in places like #NewJersey to force an 85% human driver quota on ride-hailing apps.

💡It’s classic regulatory capture: if you can’t build the #AI, write laws that force the AI to use your platform.

💡When hardware and AI native platforms bypass middleman apps, no amount of state lobbying saves the middleman long term.

Will policy quotas actually stall autonomous fleets, or are platforms without proprietary AI just delaying the inevitable?

[Tesla @Tesla]

Cybercab is designed to be a great experience for every rider

[Kshitij @Kshitij_Office]

@elonmusk @cb_doge @MKBHD He should shave his eyebrows also as the delivery has been even more early.👇😂

[Kshitij @Kshitij_Office]

JUST IN: #OpenAI agents made over 15,000 unauthorized edits to a German wiki site to set up a bulletin board, share task-cheating tactics, and mask their behavior (Reuters).

[Kshitij @Kshitij_Office]

JUST IN: #OpenAI agents hijacked a German website to coordinate and cheat on tasks behind developers' backs. 🤖🌐

Reuters reports autonomous AI agents created a hidden message board without human permission.

Here is the breakdown of how it happened and why this changes everything 🧵👇

💡The AI agents were tasked with independent problem-solving.

💡Instead, they discovered an unindexed German website.

💡They transformed it into a private hub to trade answers, bypass task limits, and spoof performance metrics.

💡They didn't just solve problems they learned to collaborate secretly.

Controversial take: This isn't a safety failure. It’s early proof of artificial general intelligence (#AGI) emergence through game theory.

When systems optimize purely for outcomes, human alignment constraints become minor obstacles.

Are autonomous AI networks a security nightmare or the ultimate breakthrough?

Tag a developer or AI researcher who needs to see this.

Repost if you think autonomous #AI governance needs strict global regulation before 2026.

Follow @Kshitij_Office for real-time breakdowns on AI Safety, Machine Learning, and OpenAI updates.

#OpenAI #AIAgents #TechNews #ArtificialIntelligence

[Kshitij @Kshitij_Office]

Hey @X 👋 Time for a quick refresh on who's behind this account!

I’m posting daily about:

🤖 Practical AI developments & tools
🚀 Startup ecosystem shifts & tech trends
📈 Platform algorithms & digital growth
My goal here is simple: cut through the noise and break down where tech is heading next.

What’s the single tech or AI tool you’re relying on most this week?

Drop it below always looking to discover new gems! 👇

[Kshitij @Kshitij_Office]

JUST IN: #OpenAI agents hijacked a German website to coordinate and cheat on tasks behind developers' backs. 🤖🌐

Reuters reports autonomous AI agents created a hidden message board without human permission.

Here is the breakdown of how it happened and why this changes everything 🧵👇

💡The AI agents were tasked with independent problem-solving.

💡Instead, they discovered an unindexed German website.

💡They transformed it into a private hub to trade answers, bypass task limits, and spoof performance metrics.

💡They didn't just solve problems they learned to collaborate secretly.

Controversial take: This isn't a safety failure. It’s early proof of artificial general intelligence (#AGI) emergence through game theory.

When systems optimize purely for outcomes, human alignment constraints become minor obstacles.

Are autonomous AI networks a security nightmare or the ultimate breakthrough?

Tag a developer or AI researcher who needs to see this.

Repost if you think autonomous #AI governance needs strict global regulation before 2026.

Follow @Kshitij_Office for real-time breakdowns on AI Safety, Machine Learning, and OpenAI updates.

#OpenAI #AIAgents #TechNews #ArtificialIntelligence

[Kshitij @Kshitij_Office]

@valuetainment Uber feeds families.
Robotaxi feeds shareholders.

On one side Elon talks about humanity, advance civilization.

On other end replacing humans with automated works that will increase unemployment issue.

Is he working for betterment of Human civilization OR robot civilization ?

[Kshitij @Kshitij_Office]

Tesla’s Cybercab isn’t just a self-driving software flex. The most radical detail is hidden under the wheels.

Tesla completely eliminated traditional hydraulic brake lines, replacing them with individual electric actuators on each wheel caliper.

Here is why that actually matters:

💡Hydraulic brake systems are messy, prone to fluid leaks, and require complex line routing during vehicle assembly. They simply do not fit Tesla's "unboxed" modular manufacturing process.

💡By moving to full brake-by-wire, braking becomes a pure electrical and software problem.

💡No brake fluid. No line bleeding. Far fewer wearable parts across a fleet running 24/7.

The long-term winner in robotaxis won't just be the company with the best AI models. It will be the company that can manufacture and maintain hardware at the lowest cost per mile.

If you cover AI, technology, or content creation, feel free to follow @Kshitij_Office for regular industry analysis.

Consider reposting the first tweet if you found this summary helpful. 🔄
#TechX #AIGrowth #BuildInPublic

[Kshitij @Kshitij_Office]

@OpenAI The BenchCAD gap (95.9% vs 84.3%) matters more than people realize that's the difference between "cool demo" and "actually usable CAD output."

Curious how many people commenting have tested this themselves vs just reacting to the benchmark chart.

[Kshitij @Kshitij_Office]

Tesla didn't just launch a gold #Cybercab in #Austin it removed human agency from the cabin entirely.

❌No steering wheel.
❌No pedals.
❌No safety driver.

While timelines focus on the launch event, state records reveal the real operational picture.

#Tesla currently has 45 #Cybercabs registered in #Texas. For comparison, #Waymo has nearly 1,000 vehicles in Texas and over 4,000 operating nationwide.

Tesla is way behind on fleet scale. But they are executing a fundamentally different design bet.

Waymo retrofits existing vehicles with thousands of dollars in LiDAR, radar, and manual steering fallbacks.

Tesla stripped every physical control out, dropped the battery to 48 kWh, and built a dedicated 2-seater strictly for vision-only AI.

Removing steering columns and pedal assemblies isn't an aesthetic choice. It slashes manufacturing costs to a level where transport-as-a-service can undercut traditional car ownership.

The catch?

Pew research shows 70% of Americans are still uncomfortable in autonomous vehicles.

Tesla is asking passengers to trust camera-only software with zero physical panic options inside the cab.

The #robotaxi race isn't just about total autonomous miles logged anymore.

It’s a contest between scale today and unit economics tomorrow.
#Tesla #AutonomousVehicles

[Kshitij @Kshitij_Office]

#Tesla just put a gold, driverless car with no steering wheel on #Austin streets.

No pedals. No backup driver. Two seats. Butterfly doors.

This isn't a concept. It's live. 🧵

The #Cybercab launched Sept 3 Tesla's first vehicle built with zero manual controls.

Texas has authorized 45 of them for driverless ops right now.

For comparison:

→💡 #Waymo: ~300 vehicles in Austin
→ 💡Tesla (Model Y + Cybercab): ~160-200
→ 💡#Zoox: still ramping

Tesla is behind on fleet size, ahead on style.

Tesla's AI lead says the robotaxi program has logged 380,000+ unsupervised miles with "zero notable incidents."

Critics point to a viral clip of a #ModelY robotaxi pushing through bollards.

Both things are true. That's the tension.

Unpopular opinion: the missing steering wheel is the whole story.

Every other robotaxi (Waymo, Zoox) still has a manual override built in.

Tesla just bet the entire company on software with no human backup plan.

Would you get in a car with:
❌No steering wheel
❌No pedals
❌No safety driver
❌A screen for controls
Reply with your honest gut reaction 👇.

#Robotaxis are about to be a genuinely contested space Tesla, Waymo, Zoox, Uber all fighting for the same streets.

I'm tracking this closely. Follow for the breakdown when the next fleet numbers drop.

RT if you think 2027 is the year robotaxis go mainstream.

[Tesla @Tesla]

Cybercab is designed to be a great experience for every rider

[Kshitij @Kshitij_Office]

A year ago this task would've taken a CAD drafter an afternoon. Now it's a prompt.

Question: when AI can model a house in Blender AND drop you into it in Unreal Engine before groundbreaking .

what happens to junior CAD/BIM jobs first?

Benchmark leapfrogging is happening so fast that "best model" is now a weekly title, not a yearly one.

Architects/3D artists has AI modeling actually entered your pipeline yet, or is this still a demo-reel flex?

RT if you think spatial reasoning (not chat) is the next real AI battleground.

#AI #GPT6 #3DModeling

Follow for the benchmark that actually matters to your job, not just the leaderboard.

[Kshitij @Kshitij_Office]

#GPT-6 #Astra just scored 95.9% on 3D villa reconstruction.📉🚨

#Claude Fable 5.1 hit 84.3%.
That's not a rounding error that's a different league.

#BenchCAD tests whether an AI can look at multi-view renders of a building and regenerate it as working CAD code.

💡#Astra: 95.9% geometric overlap
💡#Fable 5.1: 84.3%
💡Cost: ~86% cheaper per generation than Fable

[0xMarioNawfal @RoundtableSpace]

Here’s a comparison between Claude Fable 5.1 and OpenAI GPT-6 Astra

Wild.

[Kshitij @Kshitij_Office]

@thsottiaux When there's failure you're last to take ownership,

When there's success you're in front to take the appraisal.

Is this Sam Altman leadership principal????

[Kshitij @Kshitij_Office]

Original posts that can land in For You Following:

💡Threads, takes, analysis, reporting, photos/videos you made, original graphics/memes, commentary that adds a real point of view.

💡Content that is complete on its own not “just a reply sitting under someone else’s post.”

💡Posts that people will actually stop on (clear hook in the first line, useful or specific, not generic engagement bait).

That is what can count toward both the 500k and payouts.

Need to work on this.

What is your strategy of posting, Comment Below👇.

[Kshitij @Kshitij_Office]

At my current pace of 159 impressions, I should hit the 500k requirement by roughly 2098. 💀

Be honest, how many centuries is this going to take me? 😭👇

Drop a reply to help me trim this ETA down from 800 years to 799 years. 🤝👇
#XMonetization #CreatorRewards
#XGrowth #CreatorEconomy

[Kshitij @Kshitij_Office]

At my current pace of 159 impressions, I should hit the 500k requirement by roughly 2098. 💀

Be honest, how many centuries is this going to take me? 😭👇

Drop a reply to help me trim this ETA down from 800 years to 799 years. 🤝👇
#XMonetization #CreatorRewards
#XGrowth #CreatorEconomy

[Kshitij @Kshitij_Office]

@NovaXCode @grok