Sanjay

Whenever a theory appears to you as the only possible one, take this as a sign that you have neither understood the theory nor the problem

[Sanjay @ksanjay]

1/Introducing a new feature in an established product has some unique consideration in contrast to product market fit for new products. This is my framework for ‘feature’ market fit via a) Meaningful Value Add b) Cannibalization c) Competition - a 🧵...

[Sanjay @ksanjay]

https://t.co/mzlqEFG1Yy

[Praveen Neppalli @praveenTweets]

AI agents are everywhere at @Uber. It’s great to see, but the thing that keeps me up at night is how we are going to secure them. This is something that I have been thinking about for a while.

Today, our agents run 50,000+ sessions per day across thousands of endpoints. And this isn't just engineering anymore. Employees across the company use agents that read code, run commands, call internal tools, analyze data, and act on real systems.

That scale forced us to confront an important question: How do you secure agents when your security tools can't even see them?

Traditional Endpoint Detection & Response (EDR) sees the file write, but not the prompt that triggered it. It sees the network call, but not the agent's reasoning. The intent, the thing that separates malicious from benign, is invisible.

So we built Agentic Detection and Response (ADR):

• Capture the full causal chain: prompt → reasoning → tool call → outcome, across Cursor, Claude Code, Codex, and every agent our employees use.
• Triage cheaply: a fast, high-recall first pass handles the flood of benign sessions.
• Reason deeply: only suspicious events get expensive LLM analysis, enriched with source code, threat intel, and policy context.
• Red-team continuously: an offline explorer evolves hard attack variants before attackers find them.

After 10+ months in production, the results speak for themselves:

• Hundreds of credential exposures detected across 26 categories.
• Shift-left prevention blocking secrets at 97.2% precision, before they ever leave the laptop.
• Zero false positives on our enterprise benchmark, with 2-4x the F1 score of state-of-the-art baselines.
• Every attack detected on AgentDojo, the public prompt injection benchmark.

Just as valuable as the detections are the lessons from running this in production:

• The workflow is the unit of security, not the individual tool call. Attacks hide in causally-linked chains that look benign step by step.
• Credential leakage is a far more common operational issue than prompt injection.
• Approval fatigue is real: when users approve 50+ actions per session, human oversight becomes a rubber stamp.

You can't secure agents you can't observe. And nobody can solve this alone. That is why we recently joined the Open Secure AI Alliance (OSA), and why today we're taking the next step: open-sourcing ADR.

The release includes the ADR Sensor, the detection framework, and ADR-Bench, the first enterprise agentic AI security benchmark: 302 tasks derived from real production telemetry and full coverage of all 17 attack techniques across 5 tactics, so the community can rigorously evaluate their own defenses.

Code: https://t.co/tmI0oQ1F2U
Paper: https://t.co/0xbswUbiF5

The future of AI security won't be built behind closed doors. Excited to see what the community builds on it, and what we all learn together! @UberEng

[Karl Mehta @karlmehta]

Satya Nadella reveals why every company may need its own AI model: the model becomes the new company database.

"To me, a model is like the database market."

"A firm should be able to take the tacit knowledge it has and embed it inside weights in a model that they control."

"When somebody asks me how many models should there be, I'll say as many models as firms in the world."

The contrarian part: the value may not sit in one universal frontier model. It sits in each company turning its private operating knowledge into a controlled model.

[sridhar @RamaswmySridhar]

Early results from the @snowflake's coco team on GLM-5.2 vs Opus-4.7 on dbt-bench — what the trajectories actually show 🧵

[Sanjay @ksanjay]

Reflection on Human Creativity and AI: The Virtuous Loop of Knowledge Creation

I've been reflecting lately on what truly drives progress in science, technology, and our understanding of the world as someone involved in scaling agentic AI systems at work and even in personal pursuits like ongoing SLE (lupus) research. The core idea that keeps coming back to me is this: Humans remain the only genuine source of new knowledge. AI is a powerful accelerator, not a replacement. This isn't about diminishing AI's incredible potential; it's about recognizing where true creativity and error correction come from.

David Deutsch, in works like The Beginning of Infinity, argues that knowledge creation is a uniquely human (and evolutionary) process of generating bold, explanatory ideas. Karl Popper talks about this beautifully: real progress happens through imaginative conjectures followed by rigorous attempts at falsification. We're not just accumulating data; we're forward error correction agents... spotting flaws in our models of reality, proposing better explanations, and iteratively refining them. As far as we know, this open-ended, creative capacity is what sets us apart.

Large Language Models don't originate new explanatory knowledge on their own. Their primary strength is speed and scale. Just like transformed how we search and navigate the web's existing content, dramatically improving quality and accessibility, but it didn't invent the underlying facts or pages. Similarly, AI can synthesize vast literature in moments, simulate scenarios, generate candidate hypotheses for us to evaluate, and explore implications at speeds unimaginable before. What once took teams weeks or months of manual grinding can now happen in hours.

This partnership forms a virtuous loop: Humans supply the novel conjectures, creative leaps, and critical falsification that constitute genuine knowledge creation. AI then acts on that new insight: rapidly testing variations, scaling analysis, and feeding refined data or ideas back into the human process. The advantage lies in the accelerated velocity of the loop itself, enabling us to tackle complexity at new scales while preserving the integrity of discovery.

What if we removed humans from this loop entirely? Progress would likely stagnate. AI could recombine and optimize existing knowledge at blazing speeds, but without bold new conjectures and human-style forward error correction, it risks becoming self-referential, hitting diminishing returns, accumulating uncorrected blind spots, and failing to produce new explanations. The open-ended growth would dim.

Conversely, if we removed AI? Humanity would keep advancing, just very slowly. We've done so for centuries through painstaking research, experimentation, and ingenuity. The fundamental engine - human creativity and criticism - remains fully intact. But we'd forgo the velocity (which we have come to expect thanks to AI) that lets us iterate faster on complex problems, making the journey toward solutions slower and more resource-intensive.

AI makes humans faster ; Humans make AI useful - this is the most valuable virtuous loop of my lifetime.

[Satya Nadella @satyanadella]

https://t.co/vLmiBKTtX3

[Sanjay @ksanjay]

AI as a revenue accelerator ; Large Taste Layer by Spotify. This is the natural next step function move.

https://t.co/JszqshatXC

[Sanjay @ksanjay]

I’m still refining this setup and learning from my mistakes, but sandboxing feels like the only responsible way to experiment with agentic workflows on a primary machine. If you’re building or deploying local agents right now, how are you balancing power and privacy? (6/6)

[Sanjay @ksanjay]

I’ve been experimenting with Hermes Desktop on my primary MacBook. It’s an incredible tool, but running an autonomous agent natively on a machine with all your personal data is a bit terrifying. Here is the simple, sandboxed setup I’ve landed on to keep things safe. I call it the "humble mode"🧵 (1/6)

[Sanjay @ksanjay]

Fourth step: Enforce strict API hygiene. If the agent absolutely needs cloud tools to finish a task, I only supply scoped, short-lived API keys with hard spending limits. Never hand over root credentials to an autonomous system, no matter how capable it is. (5/6)

[Sanjay @ksanjay]

@Itaxh1 Appreciate it. staleness bugged me given rate at which pricing changes. Agent-fetchable docs is cool. Feel free to index it; it's all public + MIT. I'll take a look at OZ when I get a sec 🙏

[Sanjay @ksanjay]

Most multi-step LLM pipelines route every sub-task to a flagship model. It works... but you're paying Opus prices to plan a function name.
So I built a tiny open-source router that sends each sub-task to the cheapest model that can actually handle it. 🧵

[Ashwin Kumar @Itaxh1]

This is clean. The live pricing/capability registry is a smart touch — stale provider info is painful.

Would be cool to make the SDK/docs agent-fetchable too.

I’m building an MCP that indexes SDK/API docs so agents can pull the right context directly instead of relying on web search.

Happy to index simpltknopt here: https://t.co/wssGG40noX

[Sanjay @ksanjay]

@DegenApeDev https://t.co/I6jA8vdG6n

[Sanjay @ksanjay]

Most multi-step LLM pipelines route every sub-task to a flagship model. It works... but you're paying Opus prices to plan a function name.
So I built a tiny open-source router that sends each sub-task to the cheapest model that can actually handle it. 🧵

[Sanjay @ksanjay]

@DegenApeDev https://t.co/J3C9ukcRwx

[Sanjay @ksanjay]

1/Introducing a new feature in an established product has some unique consideration in contrast to product market fit for new products. This is my framework for ‘feature’ market fit via a) Meaningful Value Add b) Cannibalization c) Competition - a 🧵...

[Alexa | Startup founder @alexabelonix]

@ksanjay clean dev work.

[Compile And Push @compileandpush]

@ksanjay Nice router design. How are you handling load balancing?