
Security research, mathematics, programming | Co-Founder @DecurityHQ
Joined January 2012
- Tweets547
- Following231
- Followers1.4K
- Likes389
Omar "Beched" Ganiev retweeted
Introducing Locals: an Apple Silicon engine that runs local LLMs up to 3x faster.
It is built with a single mission: to free you from cloud providers and run frontier-class models locally without needing massive compute resources.
It’s simple. Run our engine, and your coding harnesses and autonomous agents can immediately use it as a drop-in replacement for Ollama, MLX-LM, or llama.cpp.
Under the hood is a theoretical breakthrough in complexity theory, enabling substantially faster algorithms for DeltaNet architectures, grammar-constrained decoding, tool calling, and structured outputs.
Open source and available now on GitHub.
The hack was not stopped apparently, the draining continues
🚨 @binance - Loss ~$201.2M (2026-10-29)
Network: Ethereum
Type: Delegation Abuse (EIP-7702) / Private Key Compromise
Follow-up sweep in the Binance hot-wallet drain. The wallet 0x28c6…1d60 carries an EIP-7702 delegation (code 0xef0100 + 63e842b0…), pointing at the same operator-gated wallet-drainer kit. In this tx the attacker operator EOA 0x7c8601d4 (NOT the account owner) called sweep(address recipient, address[] tokens) = 0x8a179be4, and the delegate code transferred the wallet's inventory — 200.40M USDT, 369.8 ETH, plus LINK, UNI, ENA, COMP, AAVE, QNT, PEPE, SHIB, ONDO, PENDLE, IMX, MORPHO, NIL and more — to attacker collection 0x5227f603c1ee34ce3c06efe6477003bdfccb2f8c. Net ~$201.2M in this transaction. Setting a 7702 delegation requires a signature from the account's own key, so this is a signing/private-key compromise of an exchange hot wallet, extracted via an operator-gated universal sweep (one of several sweeps totaling ~$2.5B).
TX: etherscan.io/tx/0x55ba4dabc4…
Attacker: etherscan.io/address/0x7c860…
Victim: etherscan.io/address/0x28c6c…
⏱️ Real-time alerts: defimon.xyz
defimon.xyz subscribers got a detailed llm-triaged alert at 00:27 and the raw initial alert before that in real time. 10 minutes before anyone on X knew
🚨 @binance - Loss ~$2.53B (2026-10-29)
Network: Ethereum
Type: Delegation Abuse (EIP-7702) / Private Key Compromise
The Binance hot wallet 0x28c6…1d60 was made to carry an EIP-7702 delegation (code 0xef0100 + 63e842b0…) pointing at a wallet-drainer implementation — the same kit flagged pre-drain for operator 0x7c8601d4 (Tornado-funded). In this tx the operator EOA (NOT the account owner) called sweep(address recipient, address[] tokens) = selector 0x8a179be4, and the delegate code transferred the wallet's entire token inventory — ~1.62B USDT, 167K ETH, UNI, LINK, PEPE, ENA, wBETH, xaut, PAXG, WBTC and ~60 other assets — to attacker collection address 0x5227f603c1ee34ce3c06efe6477003bdfccb2f8c. Net ~$2.53B drained. Setting a 7702 delegation requires a signature from the account's own key, so this is a signing/private-key compromise of an exchange hot wallet, extracted via an operator-gated universal sweep.
TX: etherscan.io/tx/0x53bd38b816…
Attacker: etherscan.io/address/0x7c860…
Victim: etherscan.io/address/0x28c6c…
⏱️ Real-time alerts: defimon.xyz
Omar "Beched" Ganiev retweeted
🚨 @binance - Loss ~$2.53B (2026-10-29)
Network: Ethereum
Type: Delegation Abuse (EIP-7702) / Private Key Compromise
The Binance hot wallet 0x28c6…1d60 was made to carry an EIP-7702 delegation (code 0xef0100 + 63e842b0…) pointing at a wallet-drainer implementation — the same kit flagged pre-drain for operator 0x7c8601d4 (Tornado-funded). In this tx the operator EOA (NOT the account owner) called sweep(address recipient, address[] tokens) = selector 0x8a179be4, and the delegate code transferred the wallet's entire token inventory — ~1.62B USDT, 167K ETH, UNI, LINK, PEPE, ENA, wBETH, xaut, PAXG, WBTC and ~60 other assets — to attacker collection address 0x5227f603c1ee34ce3c06efe6477003bdfccb2f8c. Net ~$2.53B drained. Setting a 7702 delegation requires a signature from the account's own key, so this is a signing/private-key compromise of an exchange hot wallet, extracted via an operator-gated universal sweep.
TX: etherscan.io/tx/0x53bd38b816…
Attacker: etherscan.io/address/0x7c860…
Victim: etherscan.io/address/0x28c6c…
⏱️ Real-time alerts: defimon.xyz
Is new phuip-fpizdam coming up?
🫡 🫡 🫡
Concentrated liquidity depends on exact delta math. @DecurityHQ audited those calculations and the calldata bounds on VM hook paths. One of 8 independent Aqua and SwapVM audits.
my honest opinion: if you as a DeFi project or hedge fund know of defimon.xyz/ and haven't subscribed, then you deserve to be the next project listed on rekt news
How to balance capabilities of security tools and phishing tools?
Have you heard about ERC-7715? It allows a dApp to request human-readable permissions from a user's smart account to act on their behalf. E.g. stream fixed amounts of tokens or revoke approvals.
So it made me thinking what if it could allow a watcher service like @DefimonAlerts to withdraw a position from your wallet if a threat is detected? Operational 24/7 on its own, without a hurry to open your wallet to submit a withdraw tx.
The only wallet that supports this kinda experimental tech is @MetaMask (new users get 7710 smart accounts by default). However MetaMask supports only a limited set of permissions that do not allow fixed calldata (which makes sense).
That's why I decided to expand the allowed permissions with `erc20-token-swap` which allows to swap your DeFi positions into safe stablecoins in the event of a protocol hack. PR: github.com/MetaMask/snap-771…
Inspired by revoke.delegate by @ayushbherwani ethglobal.com/showcase/revok…
Liquidity providers: it’s time to wake up.
Use 1inch Aqua to find more activity in more markets, without letting your tokens out of your wallet.
Risk-controlled execution meets full self-custody.
No, you aren’t dreaming.
Here’s how it works:
⬇️
Omar "Beched" Ganiev retweeted
The full story also includes the Move VM bug that put $70B at systemic risk. I guess @TheStreet skipped that part.
Aptos just hit a new quarterly high: +16M transactions in a single day.
Every transaction burns APT — 235.2K in the Last 30D, 1.4M since mainnet. Gas fees at $0.0005 since the 10x fee increase.
More usage, less APT. Read the full story via @TheStreet:
thestreet.com/crypto/innovat…
1/
we found a bug in the Aptos Move VM that put up to $70B at systemic risk. type confusion at the execution layer. a ~90% success rate across hundreds of simulated runs on a 30+ validator cluster. cost to build the attack infrastructure: $3,000.
Conducted by @kemmio , to our knowledge this is the first public research that showcases how to land a sophisticated multi-block attack in real-world environments. It includes mempool feng shui, block production specifics and about a dozen of other primitives and tricks chained to get to near-perfect exploitation results.
Nonetheless, Aptos called it "extremely low exploitability."
[hexens.io/research/aptos-hij…]
Omar "Beched" Ganiev retweeted
Replying to @Mudit__Gupta @hexens
Btw, we wrote a kick-ass article about the vulnerability. A very interesting read:
hexens.io/research/aptos-hij…
Real-time threat detection vendors like to brag that they coulda-woulda-shoulda.
We started to change this and actually save funds from thefts.
Quietly undoing the damage to the industry
Real-time DeFi monitoring is how anomalies get caught early. @DefimonAlerts is building not only an alert system, but also a white hat rescue bot that can protect defi in real time.
Learn more: youtube.com/watch?v=ioGtupYR…
Support: qf.giveth.io/project/defimon…
Defimon is one of the few real-time threat detection systems that actually impacted our industry
How we cut block latency 4x and scaled exploit detection to 8 chains: our CTO @theRaz0r shares insights on how @DefimonAlerts works under the hood on the @Quicknode blog 👇
some counterpoints:
1) arguably, llms are better at writing exploits rather than finding bugs.
If so, then given all the required bugs for the chain were discovered, it's relatively easier to automate the full chain generation
2) no need for the full chain to assess the risk level. Same llms can be used for the prioritization.
A good vuln researcher can look at a crash report and say "i think i can get this to an rce in 3 weeks". No need to actually spend those 3 weeks.
This intuition might be built within the agents as well
The market is still underrating the gap between finding bugs and actually proving exploitability.
A lot of companies argue that vulnerability discovery is becoming commoditized, that multiple models can now spot bug patterns, and that results like Anthropic Mythos are reproducible.
That misses the real technical bottleneck. Finding the bug is not the hard part anymore. In many cases, that’s close to a solved problem.
What matters is whether a model can turn that finding into a real exploit: chaining primitives, building a working test case, and demonstrating that the issue is actually exploitable without a human carrying it across the finish line.
That capability matters on both sides:
- offensively, because exploitability is what separates signal from noise
- defensively, because you need a credible proof case to build the right fix
We’re probably heading into another market correction where a lot of “AI companies” that are really just UI wrappers get washed out.
The next winners won’t be the ones with the nicest interface. They’ll be the ones with real underlying capability.
when ai usage in security is not just for likes but for actual good.
True whitehats at your service!
1/ Since inception, Decurity has focused on real-time threat prevention. Our flagship product @DefimonAlerts has prevented hacks with direct impact exceeding $20M.
But in 2025-2026, the world changed. AI is disrupting cybersecurity - and we had to adapt.
🧵
winning seconds means preventing millions in losses
⚡ Every second matters when a DeFi exploit hits onchain.
That's why we moved @DefimonAlerts from self-managed pipelines to @QuickNode Streams and went from 3 chains to 8, cut latency 4x, and freed us to focus on what actually matters - detection.
















