
Founder @Scarai_xyz. it tests security fixes against the real exploit before they ship. AI researcher (AI safety, agentic RL)
Mumbai, India
Joined March 2016
- Tweets15.7K
- Following1.6K
- Followers760
- Likes99.1K
Pinned Tweet
"I remember reading this exact bug 18 months ago in a @code4rena contest. I cannot, for the life of me find it again."
Auditors don't have a search problem. they have a recall problem.
Scar fixes that.
R@10=0.901 over 232k smart contracts at 114ms.
live demo at scarai.xyz.
The Verus bridge was drained for $11.5M in May, then again for $7.54M in July: same contract path, same bug class.
SCAR tests a security fix against the real exploit before it ships.
$20,000 is pledged. $2,500 more by 25 Nov puts SCAR in @thedaofund's 400 ETH security vote.
How the round works: TheDAO puts 400 ETH toward Ethereum security and its badge holders vote on where it goes. A project joins the vote once backers cover 25% of its budget. SCAR's is $90,000, so the line is $22,500.
Thank you Intellimation.ai and ANON HEROs for the first $20,000.
1
33
"we patched it." ok but did the patch actually work?
I ran 48 test patches against 4 real hacks and replayed the real exploit on each.
26 stopped it.
15 of those 26 were still broken. 8 left another way in, 7 broke normal use.
Scar checks this before a fix ships.
1
4
332
Scar is an approved card in @thedaofund's security round. needs $22,500 in pledges by 25 Nov to make the ballot. $10,000 is in.
pledges are only collected if the card is fully funded.
ship or review fixes? send me one, I'll run it free.
initiatives.thedao.fund/init…
You see a insanely good crypto card thats paying 25% off on AI subscriptions you get wet-> sign up-> not available in your region.
Every fucking time.
3
3
100
Excited to share that our team was accepted into the @E2B for Startups program in less than an hour! 🚀
We received $20K in E2B credits + E2B Pro Tier to help build @ScarRetreival , open evaluation infrastructure for AI smart-contract security: an agent is handed a contract that was actually exploited on-chain, writes a patch, and we replay the real attack against that patch on recorded chain state. Every episode is untrusted, model-written Solidity that has to compile, execute, and be graded by tests the agent must not read or overwrite. We have already caught one agent overwriting its own reward file and another reading the grader source from inside the container, so the isolation has to be real.
Big thanks to @E2B — handles all the sandbox infrastructure so we can focus on building the actual product. If you are early startup, check out if you are eligible for the program here: e2b.dev/startups!
Update on SCAR (AI retrieval for smart contract security), funded in the Ethereum Security QF round: $2,363 in donations + 2.21 ETH matching.
Thank you @thedaofund @quantstamp @wintermute_t for the matching pool, @Giveth @griffgreen for running the round, and everyone who donated to an open-source security project built by one person.
This update is late. That's on me. Here's where the money went, what it built, and what's next.
TheDAO's Ethereum Security QF Round matching funds have officially been distributed 🛡️
What started as a 500 ETH matching pool from @thedaofund ultimately grew to 638 ETH+ thanks to support from across the ecosystem.
134 projects have now received funding to continue their work strengthening Ethereum security.
A huge thank you to every donor, contributor, and project that made this round possible.
Last day to support @ScarRetreival at @Giveth
Every donation goes to open source harness that the security auditing needs.
qf.giveth.io/project/scar:-a…
Building the first end-to-end smart contract audit stack where every layer is grounded in retrieval at @ScarRetreival
The keystone is SCAR — a sparse code retriever I built that hits R@10 = 0.901 retrieving from a 232k-contract corpus in 114ms.
BM25 collapses to 0.308 at that scale.
What that unlocks across the stack:
Layer 1 (next) — agent harness calls into the tool layer
Layer 2 (built) — SCAR + reranker + context-aware MCP, exposed to any agent (Claude/GPT/Qwen) over MCP
Layer 3 (training data built, RL env next) — GRPO with verifiable rewards: the retriever grades the model Layer 4 (next) — Qwen3.6-27B, SFT → GRPO against SCAR
The novel bit: SAE-LoRA — low-rank adaptation of Sparse Autoencoder encoder weights. 4.6M params (0.3% of the backbone) turns reconstruction features into retrieval-discriminative ones. 37.6× over the frozen SAE baseline.
Weights, training pairs, eval set all open-source.
Building the first end-to-end smart contract audit stack where every layer is grounded in retrieval at @ScarRetreival
The keystone is SCAR — a sparse code retriever I built that hits R@10 = 0.901 retrieving from a 232k-contract corpus in 114ms.
BM25 collapses to 0.308 at that scale.
What that unlocks across the stack:
Layer 1 (next) — agent harness calls into the tool layer
Layer 2 (built) — SCAR + reranker + context-aware MCP, exposed to any agent (Claude/GPT/Qwen) over MCP
Layer 3 (training data built, RL env next) — GRPO with verifiable rewards: the retriever grades the model Layer 4 (next) — Qwen3.6-27B, SFT → GRPO against SCAR
The novel bit: SAE-LoRA — low-rank adaptation of Sparse Autoencoder encoder weights. 4.6M params (0.3% of the backbone) turns reconstruction features into retrieval-discriminative ones. 37.6× over the frozen SAE baseline.
Weights, training pairs, eval set all open-source.
"I remember reading this exact bug 18 months ago in a @code4rena contest. I cannot, for the life of me find it again."
Auditors don't have a search problem. they have a recall problem.
Scar fixes that.
R@10=0.901 over 232k smart contracts at 114ms.
live demo at scarai.xyz.
Never got the opportunity to work with C4.
I mailed them for using their data for @ScarRetreival but never got the revert back.
And I don't think thats a case now left, bummer.
Replying to @code4rena
After careful consideration, we’ve made the decision to wind down @code4rena. This community has meant a great deal to everyone who has been part of building it, and sharing this news is not easy.
2
75
Won $400 at the @lifiprotocol DeFi Mullet Hackathon — developer track. Adding $100 of my own.
The full $500 → @Giveth Ethereum Security QF round.
qf.giveth.io/qf/ethereum-sec…
$10 to each of 50 projects securing this ecosystem.
Thanks @PhilippZentner @brucexu_eth @kenny_io @arjunnchand
Replying to @brucexu_eth
Developer Tooling 获奖项目:
🥇 EarnForge —— 一个覆盖 SDK、CLI、React hooks、MCP tools、bot 和 studio 的 LI.FI Earn 全栈工具包:
xitter.cf/LifeHodlr/status/20442…
Less “someone should fund open-source security”
More “I just did” 🫡
QF math: 100 people donating $10 each beats one whale dropping $500.
500$ just got 12,920$ matching to 50 projects!
$1M+ pool. 134 projects. ~60 hours left.
If you’ve been waiting on this round… it’s now.
qf.giveth.io/qf/ethereum-sec…
1
2
81
UX is still so much broken in web3 or I'm not updated with current meta?
Ex-Classmates got excited that I'm launching my research into a product and they wanna donate for the cause but cex->p2p->wallet->verify->donate.
Onchain ramping is so much broken with absurd quotes.
122
been heads-down on this for 5 months.
The SCAR technical report is out on the project handle.
R@10 = 0.90 over 232k Ethereum contracts (BM25 collapses to 0.31). 47× faster latency. SAE-LoRA, 4.6M added params.
Paper soon on Arxiv!
Full ablations + EVMBench validation in the thread.
@ScarRetreival 👇
TECHNICAL REPORT — TL;DR
Every AI audit tool launching in 2026 has the same hidden weakness.
They reason about code without grounding in actual audit precedent. No retrieval substrate underneath.
SCAR is that substrate.
Retrieving over 232,000 Ethereum smart contracts: BM25 collapses. R@10 = 0.31
SCAR holds. R@10 = 0.90
Same hardware SCAR is 47× faster (114ms vs 5,433ms).
The missing layer underneath every audit AI👇
2
115
Replying to @brucexu_eth
Developer Tooling 获奖项目:
🥇 EarnForge —— 一个覆盖 SDK、CLI、React hooks、MCP tools、bot 和 studio 的 LI.FI Earn 全栈工具包:
xitter.cf/LifeHodlr/status/20442…
1
1
58
Heya!
I just built EarnForge with LI.FI Earn — the missing developer toolkit for 635 vaults across 16 chains.
One typed SDK powering 7 surfaces: CLI, React hooks, MCP server, Telegram bot, agent skill, docs site, and a visual studio.
18 API pitfalls handled. 458 tests. All on npm.
@lifiprotocol @kenny_io
Track: Developer Tooling
#DeFiMullet
The CLI is where it gets fun.
earnforge doctor — runs all 18 known Earn API pitfalls against any vault. Null APY rewards, string TVL, missing pagination cursors — checked in 10 seconds.
earnforge risk — 5-dimension safety score (TVL, APY stability, protocol maturity, redeemability, asset type).
earnforge suggest — give it $10K and a strategy, it allocates across chains with risk weighting. ~10% expected APY.
1
1
114
Doesn't make sense for me, you must hold it, but do you too?
Or just talks?
You want jeeters to hold start with yourself.
Set an example, buyback old airdrop tokens and hold them in portfolio that will collectively boost overall landscape.
You will get jeeters trust.
If you want to reward for holding previous drop.
Then it should be aligned both ways.
Tokens mentioned should be in the portfolio of the new team (e.g: Espresso must have arb, op, ens etc in their portfolio)
You want to reward long term, then it must be collectively of new and old projects with users.
39
If you want to reward for holding previous drop.
Then it should be aligned both ways.
Tokens mentioned should be in the portfolio of the new team (e.g: Espresso must have arb, op, ens etc in their portfolio)
You want to reward long term, then it must be collectively of new and old projects with users.
This is awesome. Espresso is launching their airdrop using holder scores (an idea I proposed last year), programmatically rewarding users who held previous airdrops in their own airdrop. First step toward a user building long-term reputation by how they behave in the airdrop.
Basically, jeeters get less tokens. This is exactly how any bookbuilding process works; it's how businesses choose their partners, how founders choose VCs, and with the help of a little data analysis, it can now be how projects can choose their initial holders.
Cool to see! Curious how this plays out in the holding data for $ESP after it launches.
(You can see from the comments that airdrop farmers hate it, which is a feature IMO)



















