AI tools like Claude, ChatGPT and DeepSeek are igniting a new arms race in the world of cyber security. It will get worse before it gets better.
Meanwhile, Sridhar Vembu terms AI spending a bubble, OpenAI and Anthropic aim to do drugs to get out of the possible bubble, and Jack wants you to talk to Dessa.This week: SBI leaked data of thousands of customers. Hundreds of banks in South America were hit by coordinated cyber attacks.
India’s Censor Board website broken into and data stolen. Quest Diagnostics, a large medical lab in the US, breached by hackers and data of thousands of patients stolen.
Cloudflare CEO Mathew Prince said his company saw millions of attacks on websites within days using recently revealed critical PHP bugs. News came in thick and fast from the world of cyber security last week.
Not all of it was due to AI. eMudhra incident, for example, was a clear case of human stupidity. But a lot was AI.
The biggest news, of course, was the PHP vulnerability. This is a technology that powers millions of websites in the world.
Human researchers using AI tools found two PHP bugs so critical it made the secretive NSA tweet out a warning. It even made the office of the Japanese Prime Minister talk geek.As harrowing as the week has been from the cyber security point of view, the consensus is this is just the beginning.
AI tools like Fable and Astra, and now possibly Kimi 3.1 and DeepSeek Flash 4.1, are so good at finding software bugs that the world is turning into a whack-a-mole game between those who attack and those who defend. Given the pace at which AI can find bugs, write exploits and let hackers do strafe runs, it is not clear if defenders will prevail.
George Kurtz, CEO of cyber security firm Crowd Strike, called it the dawn of a new era. Highlighting how AI was changing the equation, he said, “You’ve got data breaches.
You've got ransomware. You've got the ability to steal trade secrets.
AI is democratising who can do this. We're seeing many different adversary groups with very low skill, now able to conduct nation, state level attacks for, basically, the cost of tokens and a prompt.” The implications are significant, even if at the moment people like us — the normies as nerds call us — continue to live in sunshine and bliss.
But no delusions. The world is into a dangerous territory, the one where cyber security is going to be a headache so big there is a real risk of people giving up.
As the week progressed, someone put the feeling into words on X. “The internet is over as we know it,” they said. Or take the word of Mathew Prince, who tweeted a couple of days after the PHP fiasco when another critical bug was found in a key technology. “This one may be as bad or worse than the PHP ones.
Get used to lots of patching,” he wrote on X after OpenAI announced that its AI had found a bug in Brotli, a library used in technologies like Bun and Node JS. The problem with AI is that it speeds up the entire process of finding a bug, exploiting that bug, and the defences that can be put up against the exploit.
Similar to how social media and WhatsApp did for misinformation around a decade ago. The speed is so great that it defies human ability to comprehend and control it.
Currently, cyber security is proactive. That means most bugs are found and fixed before they can be exploited at a scale.
But AI turns it upside down. It forces humans into a defence because now bugs can be found fast and can be exploited faster.
This is a dangerous place for humanity to be in because it strips away the certainty we take for granted in critical services and infrastructure. It makes the entire internet mushy and fills people who use it with doubts.
It will have repercussions. Sridhar Vembu believes we are in a bubbleAlmost every week or a couple we will revisit this topic: is AI a bubble or not?
That is because there is no answer to it yet. If it is a bubble then we may have to wait for another 6 months to a year before we find out.
And if it is not, then we may have to wait another year or two before we can be certain. Currently, different people have different readings of the situation.
Sridhar Vembu, at the moment, is in the AI-is-a-bubble camp. On X he quoted a report that claimed roughly 75 per cent of all “AI spending right now comes from AI companies spending VC investment dollars”.
Then he added his own view on it. “As a rule, we have stayed away from monster-bets-on-other-people's-money approach to business,” wrote Vembu. “We are investing in foundational R&D in AI but will await a less bubblicious time to invest in infrastructure.”What should we make of it? Like many other people Vembu is sceptical of the AI boom.
Rightly so. As a technology AI seems one of the biggest breakthroughs in human history.
Its business case is yet to be firmed up. It will eventually be a critical layer on which our world will function, similar to internet and computers, but it is not certain that it will allow companies like Amazon, Microsoft, OpenAI, Google, xAI, Meta and Anthropic to make money on hundreds of billions of dollars they have poured so far into AI.
Two things can be true at the same time. Railway is one of the fundamental technologies in the world right now.
But in the mid 1800s that did not help companies and banks when almost one third of them collapsed after Railway Mania ended. And because two things can be true at the same time, I feel Vembu’s comments are also a reflection of how much capital companies can raise in India and how easy or difficult it is for an Indian tech company to play the AI game as it is unfolding currently.
For an Indian company like Zoho, which does not have the heft of Reliance or Tata, it is nearly impossible to be part of the current AI speculation. The ticket size to sit at the gambling table itself is prohibitively expensive.
You need tens of billions of dollars to enter the game where Zuckerberg and Nadella are playing. Even if they want to play, most Indian companies don’t have that kind of money.Are the Chinese cheating?After a brief lull, the Chinese AI labs are active again.
This week Moonshot came out with its latest AI model — Kimi 3.1 Pro. This is a full-fat AI model with nearly 3 trillion parameters.
Moonshot announced it with a sheet of benchmarks attached to it. The figures are spectacular.
Kimi 3.1 Pro beats or matches the likes of ChatGPT Astra and Claude Fable 5.1. This means Chinese are now almost as good as Americans in AI.
Give or take some. No more six months behind.
No more catching up in the next release cycle. They are now shoulder to shoulder with Sam Altman and Dario Amodei.
This is impressive. More so when you consider that Kimi 3.1 Pro is also going to be an open source AI model with its weights set to be released on November 4.
At the same time, there is the updated DeepSeek 4.1 Flash we talked about last week in this newsletter. As more people use it, they are finding it extremely impressive, cheap, and almost as good as large models like Claude and ChaGPT despite its relatively tiny size.But there is chatter around the new Chinese AI models, particularly Kimi 3.1 Pro.
People are finding that the model is less impressive in use than what its benchmarks suggest. It seems like an AI model which has been “benchmaxxed” that is to say it has been specifically tuned to perform well on popular AI benchmarks.
The practice is not new and is followed by almost all AI companies, American or Chinese. If true it does take a bit of shine away from everything that Chinese labs are putting out.Jack wants you to talkJack Dorsey, the nose-ringed tech maverick, wants you to talk.
No, I do not allude to BitChat, an app the Indian government hates and that lets people send messages without needing the internet. Instead, I am talking about Dessa.
Dorsey’s company Block bought Dessa, a startup by Ragavan Thurairatnam in 2020. Now using the technology and the team of that startup, Block has announced a new AI tool made for conversations.
Dessa is not the first AI with the ability to talk. But it is arguably one of the first few with the ability to talk naturally, complete with emotions, interjections, ummm and ahs. “We made the first AI model you can have a realistic conversation with that was trained entirely on speech,” Ragavan wrote on X. “Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs.
This is a GPT3 moment for speech.”Dorsey too introduced the world to Dessa with similar words. “We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person, full of emotion, nuance, pauses, and a natural cadence,” he wrote.
Is Dessa really this good? We will see when people start to get access to it.
Currently, it is invite-only. But from the demo it seems like a leap forward in human and AI conversations.
After math, drug discoveryAs I noted in this newsletter earlier, AI is now powerful enough to do theoretical math. But it is not going to pay the bill for companies like OpenAI and Anthropic.
Instead, they need to find practical use cases for their AI, the kind that impact people directly and intimately. The solved Navier Stokes does not.
Increasingly it seems AI companies like Google DeepMind, OpenAI and Anthropic believe AI could be a particularly powerful tool for drug discovery. It is good at finding patterns and drug discovery is all about finding patterns at cellular level and then figuring out ways to fix anomalies.
This week we saw two developments on this front. Anthropic announced details of its programme to find anti-venom for scorpion stings.
Anthropic recently “asked an internal research model to design minibinders (small proteins that attach tightly to a target) against scorpion neurotoxins.” Claude then found many possible candidates. “Teams at Adaptyv Bio and Twist Bioscience experimentally tested 211 designs for binding. Of these, 17 met our criteria,” the company said.
The anti-venom is not yet reaching people. But Anthropic says the process potentially makes creating one faster and that in the coming months and years it would continue to work on producing a viable drug that can help people stung by scorpions.
OpenAI, meanwhile, bought Orchestra Bio, a company founded by Ashoka Rajendra. The goal is the same: use AI to speed up drug discovery.
The technologies and people coming from Orchestra will contribute to Rosalind, the AI platform that OpenAI has created for drug discovery. “Our vision is for AI to help scientists across the full process of discovering and developing new medicines: exploring ideas, designing experiments, interpreting results… Bringing Orchestra Bio together with Rosalind is a step toward that vision,” said OpenAI. As pressure on both OpenAI and Anthropic mounts to show AI can make them money as well as can help humanity in fundamental ways, the two companies are looking at AI use within healthcare and pharmaceuticals as one of the potential goldmines.
Will it result in something tangible? Or will it just stay limited to PR and announcements?
We will see.Jensen Huang speaks to DucksOne of the coolest moments last week from the world of AI came from the stage where the Nvidia CEO was talking about the upcoming merger of AI with physical hardware. To make his point, Jensen Huang brought with him five Microducks, small bi-pedal robots that look like a cross between Pixar’s animated table lamp and Wall-E.
These ducks are toys, but somewhat smart. They waddle and quack quack in their robot voice.
For this particular demo the ducks were made smarter with an infusion of some AI into their machine brains. They followed Huang around the stage, made agreeable noises when he “spoke” to them, and followed a few commands like an eager puppy.The ducks are obviously ultra cute, thanks to their design and innocence they exhibit with their voice and child-like mannerism.
Jensen’s demo gave us a glimpse of a future where we expect to have similarly playful smart robots that would, hopefully, amuse and humour us. It would be a slice of the future different from that of super smart industrial robots, which are expected to be hulking, mean-looking, utilitarian and all focussed on work at hand.SIX STORIES YOU CAN'T MISS01Anthropic is setting up public benefit fund to prepare for the instability AI could bring in future02 Sam Altman dismisses AI job loss fears, says coding will actually make you busier03Mark Zuckerberg finally reveals what's stopping him from launching Muse AI in India04US bans agencies from using Chinese AI DeepSeek, warns sensitive data may go to China05Tech Mahindra teams up with SpaceXSI to bring Elon Musk's AI into business workflows06OpenAI resumes AI training, cuts live internet to stop agents from escaping offline world