Hope the 15 year old Vats is proud; I kept a promise and it's about to be fulfilled in a few days!

I was in fifth grade when I learned to read my parents' silence. They were talking in the kitchen, late at night, in that particular hush that parents use when they think children are asleep. I caught fragments. Bombay. Tests. Something about a doctor. The next morning,

you're only the sum of stories you can tell

How 2022-2025 has gone so far 2026 will be the same

some alumni from the NINETEEN SEVENTIES pilani batch did a concert tonight. they played eric clapton, the beatles, sinatra, some beautiful originals. in the audience were ~100 of their kids, grandkids and loved ones. many flew in from around the world. imagine being in your

used to think professionalism meant keeping clients at arm's length 2 years in and i've learned the opposite: meaningful work only happens when you actually like the people you're working with. everything else is just a grind

i think if i stop using upi i'll end up saving more

agra has got to be the wild wild west of this country ngl the stories which i've heard from folks who stayed there

Interacting with this for the last time to say that I donated most of my payout to Swabhimaan Trust in Chennai, a trust for adolescents and young adults with autism. I’ve worked with them for my graduation project and they’re doing great work to make different therapies

ppl in school were so talented, how has everyone ended up in consulting

late night gedis and rolls with the boys- life has meaning now

netflix after acquiring warner bros:

Want to actually secure your AI agents? Check out https://t.co/aQAyUvGtxU #ai #NeurIPS #llm #security #safety #aiagent #risk

We are operating autonomous systems that can access your databases, call your APIs, spend money, spend your data elsewhere, no requests, no humans. No one is expecting bad actors. This is not an exaggeration. This is just another day.

And the worst part? These are simple, almost lazy, attacks. Place malicious instructions somewhere on a page, wait for an agent to read it and watch it execute instructions it should not.

Prompt injections are real. There are examples every day. -EchoLeak compromised Microsoft Copilot with no user involvement. -94% of LLM systems are susceptible to trivial prompt injections. -Symantec exhibited the use of AI to automate credential stuffing.

Claude can access your emails. GPT can code. Gemini can search the internet. And someone just managed to make them do these things irresponsibly. For a decade we taught AI to assist users. For about six months we taught them to secure the systems.

#ai #NeurIPS #ml #llm #safety #security

We're connecting AI agents across organizational boundaries before we've solved security within a single organization. This isn't a future problem. Companies are deploying this now. In healthcare. In finance. In supply chain management. Because it works. Until it doesn't.

The worst part? Traditional security tools assume you can see everything. In federated multi-agent systems, you can't. Each organization sees its slice. Nobody has the full picture. Accountability becomes impossible.

-Message spoofing between agents. Role impersonation. Recursive prompt injection that spreads through agent networks. Coordination attacks where multiple agents collude to bypass security.