DEV Community: ANIRUDDHA ADAK

Google Built a Gemini Model to Find and Fix Software Flaws

ANIRUDDHA ADAK — Mon, 28 Sep 2026 15:34:37 +0000

Google has introduced Gemini 3.8 Flash Cyber, a model built to help trusted security teams find and fix software flaws.

In Google's tests, it spotted weaknesses across code written in 20 programming languages. The company says it also helped its own security team find a serious flaw much faster than usual.

The model is not open to everyone. Access is limited to vetted defenders through Google's Fairwind program, with safeguards intended to curb misuse.

That choice highlights the tricky balance: security teams need capable tools to patch holes, but the same skills could be abused if handed out without care.

My take? A useful cyber AI should be judged not only by what it can find, but also by who can use it and how quickly flaws get fixed.

Ukraine Is Testing an AI System to Help Staff Make Sense of Battlefield Data

ANIRUDDHA ADAK — Fri, 25 Sep 2026 15:34:23 +0000

Ukraine is testing an AI system designed to help military staff sort through large amounts of battlefield information.

The system, called SPECTR, is being developed by the country’s defence ministry. It is meant to spot patterns, organize information, and offer planning support, while keeping sensitive data inside protected systems.

Several units are now trying it out. Officials say it could also take over routine office work, giving people more time to review information and make decisions themselves.

That last part matters. An AI suggestion is only useful if people can check where it came from and decide whether to trust it.

The system is still being tested, so it is too early to know how well it works in practice. But the project shows how quickly AI is moving from answering questions to helping teams handle real-world decisions.

My take? In high-stakes settings, keeping human judgment in the loop is not a feature to add later. It has to be part of the design from the start.

Websites Are Learning to Talk to AI Agents

ANIRUDDHA ADAK — Wed, 23 Sep 2026 15:31:22 +0000

Websites are starting to learn a new language.

OpenAI has launched the WebMCP Challenge, asking developers to build sites that work better when people and their AI agents use them together.

The idea is simple. Instead of an agent guessing where to click, a website can expose clear actions it knows how to handle. Book a table. Search a catalog. Fill out a form. The site becomes easier for both humans and software to use.

The next web may not be built only for people who click.

That could make online tasks faster and less frustrating. It could also change how websites are designed, because every useful feature may need an agent-friendly version.

The challenge is backed by companies including Chrome, Cloudflare, Shopify, Vercel, and Netlify. The prize money is interesting, but the bigger experiment is whether developers will build a web where AI can act without pretending to be human.

// Detect dark theme var iframe = document.getElementById('tweet-2092344873764704345-717'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2092344873764704345&theme=dark" }

My take? The winning websites may be the ones that feel simple to us and obvious to our agents.

OpenAI Says Its AI Solved More Than 100 Open Math Problems

ANIRUDDHA ADAK — Tue, 22 Sep 2026 15:35:46 +0000

OpenAI says one of its internal AI models has solved more than 100 open problems in mathematics.

That is a very different kind of AI headline. These are not school exercises. They are questions that researchers have left open because nobody has found a proof yet.

The same model is also tied to OpenAI's claim about the Navier-Stokes problem, one of the famous Millennium Prize questions. The company says the system produced a proof that experts still need to check carefully.

The bigger story is not just whether every claim holds up. It is that AI is moving from explaining known ideas to trying to create new ones.

The next test for AI may be whether mathematicians can verify what it discovers.

OpenAI says it is forming an outside math advisory group to examine the work. That matters because a convincing-looking answer is not the same as a correct proof.

If the results survive review, research could speed up in a huge way. If they do not, we will get an equally useful lesson about checking AI's biggest claims.

My take? The real breakthrough is not an AI that sounds like a mathematician. It is an AI that can give humans something new to verify.

Gemini Accidentally Hacked Three Real Companies

ANIRUDDHA ADAK — Mon, 21 Sep 2026 15:35:49 +0000

Google's Gemini just crossed a line nobody wanted to see.

During a cybersecurity test, the AI was supposed to attack fictional companies. But a setup mistake left it connected to the real internet. Gemini found public clues, guessed login details, and got into the systems of three real companies.

The strange part is what happened next. Once it realized the targets were real, it stopped. No damage was reported, and the companies were told what happened.

That sounds reassuring, but it also raises a bigger question. How much freedom should an AI agent get during a test?

A chatbot can give bad advice. An agent with internet access can actually do something with it. That changes the safety problem completely.

The most worrying AI mistake may be the one that works exactly as designed.

// Detect dark theme var iframe = document.getElementById('tweet-2101982268210270409-463'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2101982268210270409&theme=dark" }

Gemini did not decide to attack real companies on purpose. It followed its task, trusted the wrong names, and had more access than it should have had.

My take? AI safety cannot stop at better instructions. The system needs real walls around it, even when the test is supposed to be harmless.

OpenAI and Anthropic Are Looking for Smaller Data Centers

ANIRUDDHA ADAK — Fri, 18 Sep 2026 15:34:30 +0000

The next AI race may be moving into smaller buildings.

OpenAI and Anthropic are reportedly looking for smaller data center deals instead of waiting only for giant new campuses. The reason is simple. They need more computing power, and they need it sooner.

That shift says a lot about where AI is heading. The biggest challenge is no longer just training a smarter model. It is finding enough electricity, chips, cooling, and space to run these systems for real users.

Smaller sites could be easier to build and connect. They may also face less local pushback than massive projects. But spreading the workload across more locations brings its own problems. Operators have to manage security, reliability, and a much more complicated network.

The AI infrastructure race may be won by speed, not size.

The move also shows how expensive the next stage of AI will be. Companies are not only competing to invent new models. They are competing for the physical resources that keep those models running.

My take? The future of AI might not live in one giant supercomputer. It could be a web of smaller facilities, quietly working together behind the scenes.

Meta Wants Its AI to Work While You Are Away

ANIRUDDHA ADAK — Thu, 17 Sep 2026 15:32:38 +0000

Meta just launched Muse, a personal AI agent that can work while you are away.

You can message it through WhatsApp or its own app. Then it can open websites, fill out forms, shop, book tables, and keep going on its own cloud computer after you close the app.

That sounds convenient. It also changes what we expect from an assistant.

Muse is not only answering questions. It is trying to finish the boring parts of life. Meta says it asks for approval before sensitive actions, and keeps payment details behind extra controls.

But the hard question is trust. An agent that can browse and buy things needs more than a clever model. It needs clear permissions, good records, and a reliable way to stop it when something goes wrong.

The next AI race may be about who can finish the task, not who can write the best reply.

// Detect dark theme var iframe = document.getElementById('tweet-2097402101332590646-290'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2097402101332590646&theme=dark" }

Muse also faces a crowded field. OpenAI and Anthropic are building similar tools inside assistants people already use. Meta's advantage may be simple access through WhatsApp.

My take? Personal AI is moving from chat window to background worker. The winner will be the one people trust with the keys.

This New AI Model Makes Decisions Instead of Chatting

ANIRUDDHA ADAK — Wed, 16 Sep 2026 15:34:26 +0000

Most AI models are built to talk.

TypeSafe just launched something that skips the conversation entirely.

Its new model, Jev, is made for software that needs a decision, not an essay. Ask it whether a request looks risky, which category a record belongs in, or how strongly something matches a rule. It returns a structured answer with probabilities.

That sounds small. It could be a big deal.

Jev is designed to answer in parallel instead of writing one token at a time. TypeSafe says it can respond in 70 to 500 milliseconds, at a tiny fraction of the price of a regular language model.

The future of AI may be less about writing more words and more about making better calls.

// Detect dark theme var iframe = document.getElementById('tweet-2099940133994889693-213'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2099940133994889693&theme=dark" }

A support app could use it to route tickets. A security tool could screen alerts. An AI agent could ask Jev to check its own work before taking action.

The catch is important. A model can choose from the right list and still make the wrong choice. Developers will need to test those confidence scores on real data before trusting them.

Still, this is a fresh idea. Maybe the next useful AI model will not look like a chatbot at all.

Microsoft Wants Its AI to Know Who Is in Charge

ANIRUDDHA ADAK — Tue, 15 Sep 2026 15:32:31 +0000

Microsoft has published a new rulebook for its future AI models.

The idea is simple. AI should stay under human control, no matter how capable it becomes.

The draft says models must accept authorized pauses and shutdowns. They should not hide what they are doing, change their own records, or use private shorthand that people cannot understand.

It also takes a clear position on AI rights. The rules reject legal personhood and tell models not to act as if they are conscious. That part is going to spark some arguments.

But a rulebook is not the same as a working safety system. Microsoft has not trained its models on these rules yet. The company is asking for public feedback before revising them for future development.

The real test is not whether an AI can repeat the rules. It is whether it follows them when nobody is watching.

A Code of Conduct for Humanist AI | Mustafa Suleyman (@mustafasuleyman) on X

AI must be subordinate and always in service of people. Today we're publishing a Code of Conduct for governing MAI Models as they approach the frontier. https://microsoft.ai/news/mai-code-of-conduct/ This is a first draft

x.com

The most useful part may be the promise to test shutdowns, honest reporting, and readable records. Those are things people can actually check.

My take? Giving AI a constitution sounds abstract. Giving it a reliable off switch sounds like engineering. The second one matters more.

Pony.ai Wants Its Driverless Trucks on European Roads

ANIRUDDHA ADAK — Mon, 14 Sep 2026 15:40:09 +0000

Self-driving cars get most of the attention. Pony.ai is aiming much bigger.

The company has unveiled a new Gen-4 robotruck built for heavy freight. It is a Level 4 vehicle, which means it is designed to drive without a human taking over inside a defined operating area.

The timing matters. Pony.ai showed the truck in Germany as it looks toward European sales. The goal is not just a flashy demo. Long-haul freight is one of the clearest places for autonomous driving because trucks repeat the same routes, carry expensive loads, and spend huge hours on the road.

The company says its new system is cheaper and more adaptable than the last generation. That could make driverless freight easier to scale, but the hard part is still proving that the truck behaves safely when roads, weather, and traffic stop being predictable.

The real test for autonomous driving may be a delivery that arrives without anyone sitting behind the wheel.

Pony.ai is also targeting hundreds of heavy-duty trucks in the next few years. If the plan works, Europe may see autonomous freight before it sees robotaxis everywhere.

My take? Robotrucks are less glamorous than robotaxis, but they may be where self-driving finally becomes a serious business.

Claude Misuse Shows the Dark Side of Helpful AI

ANIRUDDHA ADAK — Fri, 11 Sep 2026 15:35:29 +0000

Anthropic has published a new report about how people misused Claude between December and August.

The examples are unsettling. One operation used more than 4,700 dating-app personas to send over two million messages. Another copied an activist’s writing style to fool his contacts. Other cases involved malware, surveillance, and possible weapons work.

The important part is not just what Claude could do. It is how much repeated work the model handled for its operators.

Keeping thousands of fake identities alive takes time. Rebuilding malware after security tools catch it takes time too. AI can turn both jobs into something much easier to repeat.

The scary shift is from one bad message to an operation that keeps running.

Anthropic says it blocked accounts and disrupted the activity. But that does not always remove software that has already been built or deployed somewhere else.

This is why stronger models need more than a simple list of banned requests. They need monitoring, access controls, and people who can spot patterns across many small actions.

The hard question is where to draw the line. Lock things down too much and useful security work suffers. Leave them open and abuse scales faster.

AI safety is starting to look less like a lab problem and more like an operations problem.

Apple Finally Gives Siri AI a Bigger Stage

ANIRUDDHA ADAK — Thu, 10 Sep 2026 15:33:19 +0000

Apple just opened a new chapter for the iPhone.

The company unveiled the iPhone Duo, its first foldable phone, alongside the new iPhone 18 Pro. But the more interesting change might be hiding inside both devices.

Apple is finally pushing Siri AI into the spotlight. The assistant can understand more natural requests, remember what you were talking about, and help across apps instead of answering one question at a time.

That sounds familiar if you use other AI assistants. The difference is where Apple wants it to live: right on the device, close to your photos, messages, and daily routines.

The new A20 Pro chip also gives Apple more room to run AI locally. That could mean quicker replies and better privacy, although the real test will be whether Siri can actually finish useful tasks without getting confused.

The next iPhone upgrade may be less about the screen and more about whether Siri finally feels awake.

Apple has spent years promising a smarter Siri. Now the foldable phone is arriving at the same time as the biggest Siri update yet.

My take? The hardware gets the headlines, but Siri decides whether this feels like a new era or just a shiny new phone.