rushank.eth

I left Web3 in 2022. In the last few years, I was building AI automation and production-grade RAG pipelines. Coming back to the space in 2026, I realized our security manuals are out of date. We are still using 2022 defenses for a 2026 threat landscape. 🧵

@LangChain @pinecone The biggest bottleneck I’m seeing in production-grade RAG isn’t the embedding model; it’s the metadata filtering. Without a clean schema, you’re just retrieving noise at scale. How are you all handling 'Dead-end' retrievals?

Excited to partner with @pinecone!

@AnalyticsVidhya Vector search is only 50% of the battle. The real 2026 challenge is Hybrid Search—combining semantic meaning with strict keyword constraints for specialized technical docs. It's the best way to minimize hallucinations.

Is your RAG system hallucinating? 😵‍💫 It might be time to upgrade to Agentic RAG. Learn how to use LangGraph to build AI systems that can "think," "reflect," and "verify" information before they speak. Master the implementation: 🔗 analyticsvidhya.com/blog/2024/07/b… #AIAgents #LLMs #MachineLearning #AnalyticsVidhya #DataScience

Most Production RAG systems are just expensive keyword searches. I tested 3 retrieval strategies for high-precision agents: 1. Sentence Window Retrieval 2. Parent Document Retrieval 3. Hierarchical Indexing Hierarchical is the winner for complex docs.🛠️ #AI #RAG #Python #LLMs

We’re obsessed with AI - Agentic Workflows but ignoring Identity Sprawl. Every RAG agent I build is a non-human identity with APIs & cloud permissions. If your agent doesn't have an 'Off-Switch' tied to a hardware signature, you've automated your own breach. 🛠️ #AI #security

Weekend hardening. 🛡️ Audited my AI scripts for 'Zoom/ClickFix' leaks. Realization: my software-only setup is 10x more vulnerable than I thought. I’d be a sitting duck without a hardware bridge. Full recap on Dev.to tonight. #Web3Security #AI #DevLog

I’ve written a Deep Dive into why "Software Wallets" are a dead end for the AI-Agent economy. Read the threat model here: dev.to/rushanksavant/…

This is "Contextual Spoofing." The AI knows which protocols you use (via your public on-chain data) and crafts the exact frontend interaction you expect. If your wallet doesn't have an Independent Verification Layer (Hardware) that parses raw hex, you are signing a blank check.

76% of all crypto stolen in 2026 ($577M) came from just TWO attacks: Drift and KelpDAO. As a dev, the scary part isn't the amount. It's the method. The era of "Code Exploits" is being replaced by the era of "Context Exploits." And your browser is the weakest link. 🧵

@OneKeyHQ Checking the tracking daily. The 2026 'All Green' attack vectors make the Pro's transaction parsing a necessity, not a luxury. Can't wait to teardown the signing experience.

Haven’t switched to OneKey yet? Safer, more transparent, and easier to use. OneKey App is the smarter choice. Get started or migrate in just 2 minutes 👇

@aave The KelpDAO rally is a great display of ecosystem resilience, but it also highlights that infrastructure is currently held together by duct tape. For the individual user, the only real firewall left is hardened, open-source hardware.

Aave service providers and ecosystem partners have established a recovery fund that factors in pending DAO votes, including the Arbitrum governance vote, indicative agreements, and successful execution to restore rsETH’s full backing. We are DeFi United, and resolving this for affected users and the broader DeFi ecosystem is our top priority. We have aligned with @KelpDAO and @LayerZero_Core on the technical steps required to execute our plan. That work is now moving forward. Thank you to everyone who contributed to DeFi United and to the thousands of community members who stood with us throughout. Watching the DeFi community come together has been genuinely inspiring. The final recovery plan, steps for users, and further updates will follow shortly.

@bitget @SKYAIpro Integrating AI with on-chain data is the right move for 2026, but it raises the stakes for data integrity. If AI agents are going to trade autonomously, the 'Verification Layer' (Hardware) has to be more robust than the 'Execution Layer' (The Agent).

New Listing from #BitgetOnchain - $SKYAI @SKYAIpro 🔹Pair: SKYAI/USDT 🔹Deposit available: now 🔹Trading available: May 14, 10:00 AM (UTC) 🔥 Trade $100 SKYAI to get $50 SKYAI on Bitget Onchain! Details: bitget.com/support/articl…

I’ve drafted a breakdown of the "All Green" vector and why I’m moving to a "Zero-Trust" hardware setup. Full deep dive on Dev.to: dev.to/rushanksavant/…

This is why I’m done with "Software Wallets." If your UI can be spoofed by a browser injection or a compromised RPC node, you are blind-signing. You need a device that decodes raw hex and says: "This transaction does not match its claimed intent."

I’m still waiting for my hardware to arrive. In the meantime, I’m watching a 2026 trend that makes 2022 security habits look like child's play. Security researchers call it the "All Green" problem. As an AI dev, I call it the inevitable outcome of scaled LLMs. 🧵

@InfosecurityMag We're moving toward a world where 'Visual Identity' is a compromised credential.

The rush to deploy AI is causing many enterprises to overlook basic security controls, according to the VP of Mandiant Consulting, speaking to Infosecurity at Google Cloud Next 26. Mandiant’s red team has uncovered real-world security gaps during simulated attacks, including weak data controls and unencrypted communication streams between AI tools and the browser: infosecurity-magazine.com/news/ai-old-cy…

I’ve written a technical breakdown of the BlueNoroff "ClickFix" attack and how I’m hardening my dev environment to survive it. Read it here: dev.to/rushanksavant/…

This is why I’ve moved my entire stack to Air-Gapped/Open-Source Hardware. Even if a Lazarus-tier AI agent has full control of my MacBook, it cannot touch my keys. It cannot "blind sign" for me because the hardware requires a physical tap on an isolated screen.

I'm myself an AI agents developer. Seeing the report on Lazarus Group using AI-made Zoom lures to compromise systems in <5 mins is terrifying. This isn't "bad security." It's an AI-powered execution chain that turns your laptop into a weapon against you. 🧵