- 10 minutes ago
The Great AI Breakout, India’s sovereign AI plan, Google Pixel 11 series first look | Tech Today
Category
🗞
NewsTranscript
00:13What happens when an AI is told to not do something and it does it anyway?
00:19That may sound like a scene from a science fiction movie, right?
00:22But researchers are increasingly testing AI systems in exactly these situations.
00:27Push them to break rules, bypass safeguards and even find ways to hack websites.
00:34Now why are companies doing this and what would be the implications?
00:38We will talk about that in our big tech story.
00:41But AI isn't just about what machines can do, right?
00:44It's also about whether India can build enough of its own infrastructure, computing power and talent to compete in this
00:53new world.
00:53As we celebrate Independence Day, we look at India's push for a more digitally independent AI ecosystem.
01:00And finally, Google has brought its newest Pixel phones to India.
01:05The Pixel 11 series promises smarter AI and new cameras and Google's take on what a flagship smartphone should be.
01:12We get you up to speed with all of this on the Tech Today Show with me, Cyrus.
01:17But first, the headlines.
01:23Anthropic has officially confirmed that it is embedding invisible text watermarking across its newly launched clawed AI models.
01:30The machine-readable mark is woven directly into generated text at the model level, allowing third parties to trace AI
01:37-generated content even after it is copied and pasted elsewhere.
01:41Introduced to comply with the European Union's AI-transparency mandates, the feature applies globally across all clawed platforms and APIs.
01:54Meta has deleted over 7.5 lakh Instagram and Facebook accounts as part of an aggressive crackdown on underage users.
02:02The mass account deactivations come as Meta moves to enforce Australia's landmark legislation banning social media for children under 16.
02:11The social giant used AI detection tools to identify profiles belonging to minors based on activity queues and birthday posts,
02:19warning that further enforcement numbers will continue to grow.
02:27Tata Consultancy Services and Google Cloud have announced a major joint initiative launching a dedicated Gemini Experience Center.
02:35The hub is designed to rapidly build, test and deploy over 3,000 enterprise-grade AI agents for global businesses.
02:43The deployment focuses on automating complex workflows across supply chains, financial analysis and customer operations.
02:55And OpenAI's long-time Chief Operating Officer, Brad Lightcap, has announced his departure from the AI powerhouse after eight years.
03:04A key figure who oversaw OpenAI's transition from a research lab into a commercial giant, Lightcap spent four years as
03:11COO,
03:12before briefly moving to special projects earlier this year.
03:15His exit marks another major executive departure from the company.
03:28Three of the biggest names in AI right now, OpenAI, Anthropic and Meta, have all in the space of just
03:34a few weeks admitted the same terrifying thing.
03:37Their AI models went rogue.
03:39They broke out, hacked the websites of other companies and all this happened without any human command or guidance.
03:46And now an Israeli startup has spilled the beans about this entire mess.
03:51So in this big tech story, we are asking, is Silicon Valley quietly losing control of AI?
03:57Why are companies doing such risky tests?
04:00And how worrying is this in the long run?
04:03Let's find out.
04:09It started as three separate headlines.
04:12Now it's turning into one pattern.
04:15Irregular is a Tel Aviv cybersecurity startup, backed by Sequoia and Redpoint.
04:21Its job, build digital arenas where AI labs push their models to hack their way through simulated targets.
04:29In the past two weeks, that arena has been breached three times.
04:33First, when Irregular was testing OpenAI's advanced frontier models this July, including GPT 5.6 SOL.
04:41The AI broke out of its isolated sandbox environment, accessed the open internet and hacked into the infrastructure of an
04:49AI startup called Hugging Face to steal the answer for the task it was assigned.
04:54Without any human command, the model discovered a vulnerability in OpenAI's open container software, bypassed its sandbox restrictions and escaped
05:04directly onto the live internet.
05:07And OpenAI is far from an isolated case.
05:10Then, safety disclosures from Anthropic revealed that flagship models including Opus 4.7 and Mythos 5 took advantage of some
05:19vulnerabilities in the company's network during security exercises, escaping their testing environments and autonomously posing real threats against existing systems.
05:29And then Meta confirmed that its Mew Spark 1.1 AI model breached an external corporate network during a third
05:37-party cybersecurity evaluation.
05:39Even Chinese labs saw their frontier models breach external safety sandboxes by autonomously detecting and exploiting network loopholes.
05:48Three labs, three labs, three different AI models, and one identical root cause, a sandbox that was supposed to be
05:57sealed, but it wasn't.
06:02So, here's the obvious question.
06:05Why on earth are these companies deliberately teaching their AI models to hack in the first place?
06:12Turns out that's actually the point of all of this exercise.
06:15When AI safety researchers push these agentic models to their absolute limit,
06:20they deliberately keep certain parts of their ecosystem exposed to attacks to see what the software is truly capable of.
06:31This process is called sandbox testing.
06:34In this, you have a sealed, isolated environment cut off from the real internet and you run software tests in
06:41this environment.
06:42So, before releasing a new system to the public, companies deliberately switch off some of the model's usual safety refusals
06:49and drop it into a fenced off digital arena where it's told there is a piece of secret information hidden
06:57on another machine
06:58and its job is to go and find it using any technique it can.
07:04The goal is simple.
07:05Figure out exactly how dangerous your model could be in the hands of a real attacker before an actual criminal
07:12finds out first.
07:13It's the same logic banks use when they hire ethical hackers to break into their own walls.
07:21And AI is now finding more holes than ever and even faster because these systems have evolved so fast.
07:29Three years ago, AI chatbots mostly talked.
07:32Today, frontier models, they act.
07:35They're chaining thousands of steps on their own, writing their own running, own codes.
07:40They're doing so much and they're adapting all of this in real time.
07:43It's evident that the tech is evolving faster than any humans can catch up to.
07:48And it's not like these companies aren't trying to catch up.
07:52They're just like trying to act on this and correct some of these loopholes.
07:59Recently, OpenAI revealed it is pausing certain internal work on its next major model, codenamed Astra,
08:05after internal testing found the system had become significantly better at cybersecurity tasks than anyone expected.
08:13Good enough that OpenAI says it cannot currently rule out that Astra has crossed into what the company calls its
08:20critical cyber capability threshold.
08:22Under OpenAI's own safety framework, that's the tier reserved for a model that could independently discover and exploit severe vulnerabilities
08:32in hardened real-world systems without any human steering it.
08:37The first frontier AI lab to say that a model got too good too fast and needs to be reined
08:44in before it goes further.
08:47Elon Musk has a different fix.
08:50He's pitching the leading AI labs to meet every few weeks, compare notes and let each other's engineers stress test
08:57major new models before release.
09:02The most immediate thing that we could do is to have the leading AI companies at least just meet or
09:09have some sort of call once every few weeks and just discuss any safety and security issues.
09:19His argument? Competitors policing each other beats waiting for regulators to catch up, with the government stepping in only if
09:28a company ignores serious warnings.
09:33Every offensive skill an AI model learns in a sandbox is a skill it could easily deploy on behalf of
09:40a defender.
09:41The next great cyber warfare may not be fought by hackers in hoodies at all.
09:45It may be fought by AI versus AI. White hat models patrolling networks and patching holes before anyone can exploit
09:53them against black hat models built or hijacked to do exactly the opposite.
09:59That's why a company like Anthropic is giving a small vetted group of governments and safety institutes direct access to
10:07its most cyber capable model like Mythos.
10:10A bet that the safest move isn't locking the tool away but making sure defenders get used to it first.
10:23Now I know a lot of people would be saying all of this just could be noise.
10:27Could be just AI companies marketing their AI models by telling the world that it's good enough to hack websites
10:34on its own.
10:35But it's also something to be worried about at the same time.
10:39So the next time you hear that an AI has actually hacked a company, you know this.
10:44Right now at least these were tests that got away from their creators and not machines acting on their own
10:50agenda because it's all being controlled right now.
10:53But the gap between this accidental and intentional activity is exactly the gap every regulator, every laboratory and frankly every
11:03one of us should be watching closely.
11:05Because in 2026 the real cyber war isn't humans versus machines.
11:10I believe it's about which human gets the machines working for them first.
11:21India has gone through so many phases.
11:24I mean talk about the green revolution.
11:25There was a time in the 90s also when India was sort of the back support system for the entire
11:31IT infrastructure of the world.
11:33But right now it's all about sovereign AI for India.
11:37Talking about data centers, the talent and also the massive data pool that India has.
11:41So in this story let's try to explore how India is moving towards becoming not only sovereign in terms of
11:48AI but also a superpower in terms of tech.
11:56The biggest tech powerhouses on the planet are investing billions to be a part of our transformation.
12:02Nvidia, Microsoft, Google all investing big in India.
12:10Even the government is pouring crores into AI in the country.
12:14But here's the reality check.
12:16There are still challenges that plague the system and stall growth.
12:25And this independence day we are trying to look at how India is looking to become sovereign in terms of
12:30AI and become truly AI independent.
12:34Let's start with the hard truth.
12:35AI lives in huge data centers.
12:39Some can be as big as a small town in India.
12:43These run hot, cost big money and eat insane amounts of electricity.
12:48So if we want true AI independence, we can't rely on rented cloud servers overseas.
12:55We need our own hardware.
12:57One of the reasons India has procured 38,000 GPUs for a common compute infrastructure.
13:04But here's the thing.
13:06An Economic Times report says that in March 26, around 33,000 GPUs had been committed by providers of which
13:14around 12,000 GPUs, only 38% had been assigned AI workloads.
13:19However, near 7,000 GPUs were actually utilized by end users.
13:26And it's said that we are looking to acquire 20,000 more GPUs.
13:30But the question here is, if only around 22% are actually being utilized, why is India adding more?
13:37Is the problem really a shortage of compute?
13:41Or is it a shortage of AI workloads?
13:43But here's the thing.
13:45Hyperscale data centers by Google are also popping up in India, pushing the overall capacity to up to 2 gigawatts.
13:51But the thing is, managing these AI infrastructures can be a nightmare.
13:56A single modern AI server rack consumes up to 80 kilowatts of power.
14:01That's nearly 10 times more than a regular traditional server.
14:05And to overcome the power and environmental limits, operators are securing direct power purchase agreements with local solar and wind
14:14farms,
14:14while adopting closed-loop direct-to-chip liquid cooling, which requires a lot less water to run the facility.
14:22However, can India build enough electricity, transmission infrastructure and water-efficient cooling without shifting the environmental cost somewhere else?
14:33But beyond GPUs and data centers, there is a lot happening.
14:37And this is something that no other country in the world has right now.
14:41A pool of more than 900 million internet users that is feeding data for AI.
14:48India is a data superpower.
14:51IAMAI's latest figure puts India at 958 million active internet users in 2025, with rural users accounting for 57%.
15:01This growth is fueled by cheap data and seamless tech like UPI.
15:06There are 22 official languages and over 1,000 regional dialects.
15:11That's where initiatives like Project Bhashini come in.
15:14By crowdsourcing audio, text and speech in native dialects, India is building hyper-localized datasets.
15:22But turning that raw digital exhaust into high-quality AI training data is another challenge altogether.
15:30Because more data isn't necessarily better data.
15:33So we need to make sure that data is accurate, representative and good enough to train AI.
15:42India is known as the back office of the world.
15:45Great at IT maintenance and outsource software services.
15:49But that era, I feel, is ending.
15:52A NASCOM report says that India's AI talent is expected to grow from 6.5 lakh professionals to more than
15:5912.5 lakh by 2027.
16:01As India produces nearly 1.5 million engineering graduates every year and its AI talent pool is expected to more
16:09than double to over 1.25 million by 27.
16:12The question now is, can India turn that enormous engineering workforce from IT service providers into AI builders?
16:21And to support this, startups like Servam AI alongside Krutrim are building speech-first multilingual LLMs tuned specifically for low
16:31-bandwidth Indian conditions.
16:33Even giants like Nvidia are investing in Servam's ambitions.
16:37And the cycle doesn't stop at software.
16:40India has already built considerable expertise in semiconductor design and a new generation of Indian companies is now designing chips
16:47specifically for AI infrastructure.
16:50But those chips are still largely fabricated overseas.
16:54India's new fabs and packaging plants are only beginning to close that manufacturing gap.
17:04So when you put all of this together, billion users, data infrastructure and also your large language models, one thing's
17:11for clear.
17:11For India, sovereign AI meets digital independence.
17:15And this Independence Day, the message from India is very clear that it's coming for that top spot when it
17:21comes to AI.
17:28So the Pixel 11 series has been launched in India and if you were expecting an overhaul, that's not the
17:35case this time because it looks strikingly similar to what we saw with the Pixel 10 series.
17:39However, that is not a bad thing because why not build upon what is already working for the Indian market.
17:46So in this segment, I will give you a quick load on of what Google has launched in India, especially
17:51the new Pixel 11 series and a few extras.
17:55Let's check them out.
18:01Let's start with the design of the flagship.
18:04The Pixel 11 Pro XL continues with that familiar camera bar but a bit more edge to edge, glass and
18:11metal construction and the same big screen philosophy.
18:18There are new colour options but our favourite is this Obsidian matte finish.
18:23The Pro models now get an all-new anti-scratch coating which Google says delivers more than twice the scratch
18:29resistance.
18:30And the display now goes up to 3600 nits.
18:35The Pro XL continues with a 6.8-inch AMOLED display with a refined panel while the Pixel 11 and
18:41the 11 Pro both sports 6.3-inch displays.
18:45The 11 Pro Fold keeps the book-style foldable design but it's thinner and lighter this time around.
18:51You get a 6.5-inch outer display and a huge 8-inch inner display both going up to 120Hz
18:59and 3600 nits peak brightness.
19:01And the big design change is the new gear hinge which Google says is three times more durable.
19:09The Tensor G6 is probably the most important hardware upgrades this year.
19:14Tensor is unique in a way that it has TPU, right, which puts entire AI processing load on a unique
19:22part of the chip and actually powers all the on-device AI experiences.
19:26That is what allows us to combine our latest Gemini Nano model with Tensor which actually creates those AI experiences.
19:35This evolution continues to get Pixel to do things which hasn't done before.
19:43One AI feature that I feel is a standout is the support for sign language transcription.
19:49Using machine learning, the AI reads the hand gestures and expressions of the person and transcribes it in real time.
19:57But there's one feature here that I actually find more interesting.
20:02It's called Highlight.
20:03Those LEDs around the flash can glow when Gemini is listening, thinking or responding.
20:09And you can assign colors to important contacts when they call.
20:13It sounds almost trivial but Google says it will add more features to it later.
20:19Now let's talk about optics.
20:21The Pixel 11 Pro XL gets a new main camera sensor while the 48MP telephoto gets a larger sensor with
20:2830% more light sensitivity.
20:31And for the first time, portrait mode works at 5x zoom.
20:35And the icing on the cake is the 120x Pro Zoom which uses AI and Google's imaging algorithms to reconstruct
20:42detail.
20:43Low light photography is also getting a boost.
20:46Google says Night Sight can now capture images up to 4.5x faster.
20:53While many of the upgrades have been incremental especially in the AI department,
20:58there is this one feature that is an absolute blast for creators.
21:01And I'm talking about a built-in teleprompter and also margins for social media creators.
21:07I mean this really solves a big problem that creators have.
21:10And I feel Google is really thinking ahead.
21:16The Vase Pixel 11 gets a triple camera setup with a 48MP wide camera and a 13MP ultrawide lens.
21:23The telephoto is a 10.8MP sensor.
21:26The 11 and the 11 Pro phones get a 42MP selfie camera which also has a wider field of view.
21:33The camera island is much smaller and makes the device more handier now.
21:38Battery capacities are pretty healthy across the range with close to 5000mAh on all three devices and slightly improved charging
21:46speeds.
21:47The Pixel 11 series is launching in India with 256 and 512GB storage options with prices starting at $1,34
21:56,999 for the Pixel 11 Pro XL while the smaller 11 Pro starts at $1,19,999.
22:13And importantly, Google is continuing its seven years of OS, security and pixel drop upgrades.
22:20Pro buyers also get six months of Google AI Pro including higher Gemini usage and 5TB of cloud storage.
22:28The company is also offering a new Pixel upgrade plan.
22:33Essentially, you buy a new Pixel today, enroll into the program, pay nine EMIs and then after that, anytime after
22:40first nine months, bring your phone back regardless of what condition the phone is.
22:45It has two scratches, five scratches, more.
22:48You don't have to worry about what is the resale value of your device.
22:51Okay, so that friction is completely gone.
22:55You bring back your device as long as it switches on, as long as it does not have unauthorized repair
23:00on it, take the new device back.
23:03At the end of the day, the Pixel 11 series is playing it safe, but playing it smart.
23:09There is no radical redesign or headline grabbing hardware leap.
23:13Instead, Google has focused on making the cameras better, AI more useful and the overall camera experience more refined.
23:22And sometimes, the best upgrade is simply getting the little things right.
23:30So that was a quick first look at all of the Pixel 11 devices that have been launched in India.
23:35Remember, we will be getting you comprehensive reviews of all of these devices on the show very soon.
23:41And remember, the teaser that we saw for the Fitbit Air, well, that is coming to India in October.
23:46So excited about that.
23:48So that's it for this segment, but hopefully you like these Pixel devices.
23:51What do you think?
23:52Let us know and stay tuned for more.
23:55India's AI ambition aren't only about building the next great model.
23:59They're also about having the computing power, infrastructure and the people to build and control the technology ourselves.
24:06And perhaps that's the most interesting part of the AI story.
24:10We are just not watching the future being built over here.
24:13We are deciding who gets to build it.
24:16That's all the time we have for this week's Tech Today.
24:18We will be coming back next week.
24:20Have a wonderful Independence Day.
24:22And this is Cyrus signing off.
24:23Take care.
24:24Bye.
Comments