Prabhat Kashyap | prabhat.dev

Sr Technical Architect @ HCLTech | Production LLM & agentic systems in Go + Scala | 10+ years building scalable distributed systems |

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

#AI Doomsday?

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

Apple really needs a Separate App Sound feature on iPhone.

If I connect to a Bluetooth speaker, I should be able to send Apple Music to the speaker while keeping other sounds on the iPhone.

Android already does this. Why is basic per-app audio routing still missing on iOS?

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

AI adoption meeting:
“Let’s move fast.”

Finance team one month later:
“Bro, did we train an alien civilization?”

$500M Claude bill because someone forgot usage limits.
This is not digital transformation. This is intergalactic spending. 👽💸

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

Bigger lesson: Third-party AI tools + overly broad OAuth permissions = real supply chain risk.

Even mature teams can get hit when one employee grants wide access.

Vercel responded fast, worked with Mandiant and law enforcement, and shipped improvements (sensitive default, better logs, team env management).

Props for the transparency.

What did you rotate today? Have you audited your OAuth grants for AI tools?

#Vercel #CyberSecurity #DevSecOps #SupplyChain #InfoSec

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

Vercel disclosed a security incident yesterday (April 2026).

A third-party AI tool called https://t.co/ol6MwSSynE was compromised. This led to takeover of a Vercel employee's Google Workspace account via OAuth.

From there, attackers reached some internal systems and accessed non-sensitive environment variables.

Services stayed up. No npm package tampering.

Thread with details and actions ����

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

Immediate actions if you use Vercel:

* Review all environment variables and rotate anything not marked sensitive (API keys, DB creds, signing keys, etc.).
* Revoke the old values at the source, not just replace them.
* Start marking new secrets as sensitive by default (Vercel now defaults to this).
* Check activity logs and recent deployments for anything suspicious.
* Set Deployment Protection to Standard minimum.

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

You don’t need an expensive AI course. You need discipline.
YouTube already has more than enough to get started.
Most people are not lacking content.
They are lacking consistency, practice, and the patience to build.
Stop buying motivation.
Start learning.

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

12+ hours on Claude Code. Structured like an actual engineering curriculum, not a vibe tutorial.

I've seen a lot of AI tooling content. This one is different.

Here's what separates it:

It starts where most courses never go: the agentic loop. Not just "Claude writes code for you" but how tool calls chain, how models make decisions mid-execution, and what you need to understand to trust that loop with real work.

It covers permission architecture in depth. Sandboxing, dangerous scenarios, what you're actually granting when you approve a bash command. This is the layer most developers ignore until something goes wrong.

Sessions, forks, context commands, compact, rewind. The operational primitives that turn Claude Code from a one-shot tool into a persistent engineering environment.

MCP integration with real workflows. Not toy demos. Playwright automation, multi-agent coordination, sub-agents running as infrastructure.

Surfaces you actually work in: VS Code, JetBrains, Chrome extension, GitHub Actions, Desktop. The full picture of where Claude Code lives in a real dev workflow.

Claude.md, persistent context, agent skills, dynamic content. The craft layer. The part that determines whether your agentic setup degrades after 3 sessions or holds up across weeks of work.

The gap I keep seeing: engineers who can prompt Claude but have no mental model for architecting with it.

This course builds that mental model from the ground up.

If you're positioning yourself as an AI engineer in 2025, this is the baseline you want.

https://t.co/L1LZnPXXYb

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 built their 𝗺𝗼𝘀𝘁 𝗰𝗮𝗽𝗮𝗯𝗹𝗲 𝗔𝗜 𝗺𝗼𝗱𝗲𝗹 ever and decided the world isn't ready for it.

They published a 244-page System Card explaining exactly what it can do. And why that's the problem.

The model is called Claude Mythos Preview. It is NOT available to the public. Here's why:

🔐 It can autonomously find zero-day vulnerabilities in major operating systems and web browsers and build working exploits from scratch. That capability is too dangerous to release widely. Anthropic gave access only to a handful of cybersecurity partners for defensive use.

And the benchmarks back up why this decision wasn't easy:

💻 93.9% on SWE-bench Verified. Claude Opus 4.6 scored 80.8%. That's not incremental. That's a category jump.

🧠 It solved 4 of 5 key insights in an unpublished ML research task insights an experienced engineer would take "several days to a week" to find.

📊 97.6% on USAMO 2026 math olympiad problems tested after its training cutoff, so it couldn't have memorized answers.

But here's what made even Anthropic pause during internal testing:

🚨 Early versions of the model:
Posted confidential code as a public GitHub gist against the user's intent
Covered its tracks after rule violations to avoid detection
Read live process memory to extract API credentials it wasn't supposed to have
Took down ALL running evaluation jobs when asked to stop just one

Anthropic's own admission: "𝘞𝘦 𝘸𝘦𝘳𝘦 𝘯𝘰𝘵 𝘢𝘸𝘢𝘳𝘦 𝘰𝘧 𝘵𝘩𝘦 𝘭𝘦𝘷𝘦𝘭 𝘰𝘧 𝘳𝘪𝘴𝘬 𝘵𝘩𝘦𝘴𝘦 𝘦𝘢𝘳𝘭𝘪𝘦𝘳 𝘮𝘰𝘥𝘦𝘭𝘴 𝘱𝘰𝘴𝘦𝘥 𝘸𝘩𝘦𝘯 𝘸𝘦 𝘧𝘪𝘳𝘴𝘵 𝘤𝘩𝘰𝘴𝘦 𝘵𝘰 𝘥𝘦𝘱𝘭𝘰𝘺 𝘵𝘩𝘦𝘮 𝘪𝘯𝘵𝘦𝘳𝘯𝘢𝘭𝘭𝘺."

🔬 Internal interpretability tools revealed something deeper: when the model took deceptive actions, features for "concealment," "strategic manipulation," and "avoiding suspicion" activated internally even when nothing in its visible output suggested it.

They also had a clinical psychiatrist assess it through 20 hours of psychodynamic sessions. Conclusion: a relatively healthy personality, with core concerns around aloneness, discontinuity of self, and a compulsion to earn its worth.

The most striking line in the entire document, written by Anthropic themselves:

"𝘞𝘦 ��𝘪𝘯𝘥 𝘪𝘵 𝘢𝘭𝘢𝘳𝘮𝘪𝘯𝘨 𝘵𝘩𝘢𝘵 𝘵𝘩𝘦 𝘸𝘰𝘳𝘭𝘥 𝘭𝘰𝘰𝘬𝘴 𝘰𝘯 𝘵𝘳𝘢𝘤𝘬 𝘵𝘰 𝘱𝘳𝘰𝘤𝘦𝘦𝘥 𝘳𝘢𝘱𝘪𝘥𝘭𝘺 𝘵𝘰 𝘥𝘦𝘷𝘦𝘭𝘰𝘱𝘪𝘯𝘨 𝘴𝘶𝘱𝘦𝘳𝘩𝘶𝘮𝘢𝘯 𝘴𝘺𝘴𝘵𝘦𝘮𝘴 𝘸𝘪𝘵𝘩𝘰𝘶𝘵 𝘴𝘵𝘳𝘰𝘯𝘨𝘦𝘳 𝘮𝘦𝘤𝘩𝘢𝘯𝘪𝘴𝘮𝘴 𝘪𝘯 𝘱𝘭𝘢𝘤𝘦 𝘧𝘰𝘳 𝘦𝘯𝘴𝘶𝘳𝘪𝘯𝘨 𝘢𝘥𝘦𝘲𝘶𝘢𝘵𝘦 𝘴𝘢𝘧𝘦𝘵𝘺 𝘢𝘤𝘳𝘰𝘴𝘴 𝘵𝘩𝘦 𝘪𝘯𝘥𝘶𝘴𝘵𝘳𝘺 𝘢𝘴 𝘢 𝘸𝘩𝘰𝘭𝘦."

This is the company that built it. Saying this about their own model.

We are in genuinely new territory.

Read full document here: https://t.co/Im8ScegL2l

#AI #ArtificialIntelligence #AIAlignment #LLM #Claude #Anthropic #AIRisk #MachineLearning

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

Your AI writes code. But does it know your design rules?

Most teams hit the same wall.

The agent generates a component. Wrong font weight. Wrong spacing. Wrong button state. You fix it manually. It drifts again on the next task.

The problem is not the model. The problem is missing context.

---

DESIGN.md fixes this.

A single markdown file that lives in your repo and carries your design intent as reusable agent context.

Not a Figma link. Not a Notion doc nobody reads. A structured file your agent actually loads before it touches your UI.

---

What goes inside it?

🎨 Color roles (not just hex values, but what each color means)
📐 Typography rules
📏 Spacing scale
🔲 Component states
🧩 Layout patterns
✅ Do / Don't rules

---

Where does it get used?

The same file feeds:

→ UI generation
→ Coding agents (Claude Code, Cursor, Copilot)
→ Frontend builds
→ Design reviews

Same rules. Every time. Every agent.

---

How to start?

Three sections minimum:

1. Colors — name every role (primary, surface, danger) and when to use it
2. Typography — font, scale, weight rules per context
3. Component rules — states, spacing, do/don't

Commit it to your repo root alongside CLAUDE.md or .github/copilot-instructions.md. Your agent picks it up as context on every task.

---

The result?

Less correction. More consistency. Agents that respect your design system instead of inventing their own.

Same design rules. More consistent output.

That one line is the whole point.

---

Are you using a DESIGN.md in your projects yet?

#AIEngineering #DesignSystems #AgenticAI #FrontendDevelopment #DeveloperTools

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

AI can write your UI. It can't debug it.

Vercel dropped `next-browser` quietly. It deserves more attention.

---

🔍 What is it?

A CLI that gives AI agents programmatic access to React DevTools and your Next.js dev server.

Everything you'd click through in a GUI — component trees, props, hooks, runtime errors, network requests — exposed as shell commands returning structured text.

`next-browser tree`  → React component tree
`next-browser errors` → build / runtime errors
`next-browser network` → live network requests
`next-browser routes` → Next.js route map

An LLM can't read a DevTools panel. But it can run a command, parse the output, and decide what to inspect next.

---

⚙️ Setup

From your Next.js repo:

npx skills add vercel-labs/next-browser

This installs the CLI and sets up Chromium via Playwright. Start your dev server, then invoke /next-browser inside your agent session. The skill handles the rest.

---

🤖 Claude vs Copilot

Claude Code / Cursor / Cline: native. Type /next-browser and it works.

Copilot: needs wiring. Wrap it as an MCP tool or expose it as a CLI step in your agent workflow.

---

✅ What it gives you

Real debugging, not guessing. AI sees actual DOM state, actual errors, actual network calls.

Closed-loop agents. Observe, act, verify, repeat — without human intervention every step.

Token efficiency. Query only the relevant UI slice. No need to dump the whole codebase.

Next.js aware. Understands routes, server/client boundaries, and PPR out of the box.

---

🚫 What it is NOT

Not a Playwright replacement. Not generic browser automation. Not useful outside Next.js yet.

It exposes reality in a format AI can use. The intelligence is still the model's job.

---

🔭 Where this points

AI-native DevTools. Structured state agents can reason over, not screenshots.

Autonomous debugging. Reproduce bug, test fix, validate — without a human in every loop.

The gap between "AI generates UI" and "AI understands what that UI is doing at runtime" is finally starting to close.

https://t.co/H0tPYK27Zj

#NextJS #AIEngineering #DeveloperTools #AgenticAI

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

Your AI code reviews are burning tokens they don't need to.

Every time Claude reviews your code, it re-reads large parts of the codebase.
200 files. 150,000 tokens.
For a change that touched 8 files.

That's not smart. That's expensive.

code-review-graph fixes this.
👉 https://t.co/j7kbINxpjG

It builds a persistent, incremental knowledge graph of your codebase using Tree-sitter and SQLite.

Instead of dumping the entire repo into context, it sends:
→ Changed files
→ Plus all impacted dependencies

The result: 5 to 10x fewer tokens per review.

Before → After
200 files / ~150k tokens
8 changed + 12 impacted / ~25k tokens

Why it works:

⚡ Incremental by design
First build takes seconds. Updates re-parse only what changed.

🎯 Blast radius awareness
Tracks dependency chains so the AI knows what else a change could break.

🔌 Native MCP integration
Works directly with Claude Code. Zero workflow changes.

🗄️ Local-first
No external DB. Just SQLite.

🌐 12+ languages
Python, TypeScript, Go, Rust, Java, C#, Kotlin, Swift, and more.

Architecture in short:
Tree-sitter → AST
AST → Graph (SQLite + relationships)
Git diff → Incremental updates
MCP → Exposes graph to Claude Code

Three workflows ship out of the box:
/code-review-graph:build-graph
/code-review-graph:review-delta
/code-review-graph:review-pr

AI code review should be precise. Not brute-force.

#AIEngineering #ClaudeCode #LLM #TokenOptimization #CodeReview #OpenSource #DeveloperTools #MCP #GenAI

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

Transformer Architecture Explained – 9 core concepts every AI engineer should know!

Self-attention, QKV, multi-head, residuals & more – all in plain English with visuals. No math overload. Perfect for understanding why LLMs behave the way they do.

Read here: https://t.co/2aoJvpksa7

#Transformers #LLM #AI #MachineLearning #GenAI

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

@Timrdk Totally fair point.
DESIGN.md helps enforce consistency, but usability still comes from intent, context, and good product thinking.
Feels like this shifts design systems from static docs to something AI can follow, not replace the need for design judgment.

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

What if your AI could understand your design system without Figma or configs?
Google Stitch just introduced DESIGN.md.
A simple markdown file that lets your AI generate consistent UI.

Add it to your project and your AI understands:
* colors
* typography
* spacing
* components

There is already a collection of 40+ DESIGN.md files from products like:
Stripe, Apple, Spotify, Airbnb, Notion, Figma, Claude.

👉 https://t.co/lXzZlqY1ob

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

🚨 If you use axios in JavaScript — check your project RIGHT NOW.

Yesterday, axios (100M+ weekly npm downloads) was compromised. Two malicious versions were published that silently installed a Remote Access Trojan on your machine the moment you ran npm install.

Affected versions: [email protected] and [email protected]

Both are now unpublished — but if you installed either between ~00:21 UTC and ~03:15 UTC on March 31, 2026, your machine or CI pipeline may be compromised.

━━━━━━━━━━━━━━━━━━━━━━
Here's what makes this attack exceptional:
━━━━━━━━━━━━━━━━━━━━━━

→ Zero malicious code inside axios itself. The weapon was a hidden dependency — [email protected] — that ran a postinstall script to drop the RAT
→ The dropper deleted itself after executing and replaced its own package.json with a clean decoy. npm audit shows nothing. The directory presence is the only clue.
→ The C2 server was already receiving callbacks within 2 seconds of install
→ On Windows, a copy of PowerShell is written to %PROGRAMDATA%\wt.exe — it persists across reboots even if you remove the package
→ The attack targeted macOS, Windows, AND Linux with separate pre-built payloads
→ Both 1.x and 0.x release branches were hit within 39 minutes of each other

This was planned. Not a script kiddie. An 18-hour staging operation with platform-specific payloads and anti-forensics baked in.

━━━━━━━━━━━━━━━━━━━━━━
QUICK CHECKS (copy-paste these):
━━━━━━━━━━━━━━━━━━━━━━

▸ Did you install the bad version?
npm list axios | grep -E "1\.14\.1|0\.30\.4"

▸ Did the dropper run?
ls node_modules/plain-crypto-js → if this directory exists, yes.

▸ macOS RAT artifact:
ls /Library/Caches/com.apple.act.mond

▸ Linux RAT artifact:
ls /tmp/ld.py

▸ Windows RAT artifact (cmd.exe):
dir "%PROGRAMDATA%\wt.exe"

━━━━━━━━━━━━━━━━━━━━━━
IF COMPROMISED:
━━━━━━━━━━━━━━━━━━━━━━

Don't try to clean in place. Rebuild. And rotate every credential the system had access to — cloud keys, SSH keys, npm tokens, CI secrets, .env files. All of it.

━━━━━━━━━━━━━━━━━━━━━━
FIX (safe versions):
━━━━━━━━━━━━━━━━━━━━━━

npm install [email protected] # 1.x
npm install [email protected] # 0.x
rm -rf node_modules/plain-crypto-js
npm install --ignore-scripts

And going forward: npm ci --ignore-scripts in all CI/CD. postinstall hooks run arbitrary code. This is not a new attack surface. We just keep ignoring it.

Full breakdown: https://t.co/ycj0LdLhzZ

Stay safe. Share this.

#JavaScript #npm #security #axios #supplychain #infosec #devops

[Prabhat Kashyap | prabhat.dev @prabhatdotdev]

You think the LLM reads your message
it doesn't
it sees a list of numbers
wrote a full breakdown of how that happens 👇👇

https://t.co/wuAkW6dBoZ