Nethny

building in Solana ecosystem
exploring blockchain × AI
learning in public 🔭

[Nethny @nethny_dev]

$1.8B stolen across Bybit, Step Finance, and Drift - and not a single one was a smart contract bug.

Drift's $285M wasn't a program exploit. It was social engineering against the Security Council multisig. Durable nonces kept signatures alive while signers got compromised one by one.

We keep auditing contracts. Attackers keep walking around them.

The unsolved problem in DeFi security isn't code - it's proving that a human actually looked at what they signed.

[Nethny @nethny_dev]

@anza_xyz so Alpenglow depends on this for on-chain PoP verification - does that mean the voting program migration to BPF is blocking Alpenglow, or can they run in parallel?

[Nethny @nethny_dev]

@mert the shannon analogy is strong but it cuts both ways - fiber optics needed pure glass to work. right now the chain is pure but the endpoints aren't. every major hack this year was a human, not a contract

[Nethny @nethny_dev]

@0xYankee 52 real buyers in $10.4M of volume. fifty two. and people wonder why their organic score is low

[Nethny @nethny_dev]

inspired by @karpathy's workflow - adapted for crypto dev where half the docs are scattered across 5 platforms and outdated by next week

[Nethny @nethny_dev]

karpathy just shared his setup for LLM-compiled knowledge bases - raw docs go into a wiki maintained by agents, then you query it like a research assistant

now think about this for crypto. every protocol has docs scattered across gitbooks, github, medium, discord. half of it outdated by the time you find it

one local wiki where you ask "cheapest way to batch token transfers on Solana" and your agent cross-references frameworks, token extensions, and RPC limits in seconds

the real AI × crypto build isn't another token. it's tooling that lets you actually navigate this ecosystem

[Nethny @nethny_dev]

@ivan_nomadz @solana solid map but missing a dev tooling section. built for MONOLITH with quasar + claude code + awesome-solana-ai + trident for fuzzing - would be a solid addition to v3

[Nethny @nethny_dev]

@mert The scariest part is how low-tech these attacks are. No zero-days, no flash loans - just patience and social engineering against the people holding the keys.

We've gotten good at making contracts secure. Now we need the same rigor for everything around them.

[Nethny @nethny_dev]

The speed of the ecosystem response yesterday showed exactly how strong this community is. That coordination matters.

One thing this incident made clear: the contracts held. The vulnerability was at the human and operational layer. As the Foundation builds out those support channels, signer opsec and key management standards could be the highest-ROI investment for the whole ecosystem.

[Nethny @nethny_dev]

Agents monitoring is one layer, but the core issue is the moment of signing. An agent can flag a suspicious tx - but can it verify that the human signer actually understood what they approved?

That gap between detection and deliberate human decision is where these attacks keep landing.

[Nethny @nethny_dev]

For context:
- Bybit ($1.5B, Feb 2025): compromised Safe wallet dev, UI manipulation + blind signing
- Step Finance ($40M, Jan 2026): executive devices compromised, project shut down
- Drift ($285M, Apr 2026): Security Council multisig social engineered via durable nonces

@vibhu confirmed: "no program or smart contract exploit"

The attack vector has moved. The defensive stack hasn't.

[Nethny @nethny_dev]

$1.8B stolen across Bybit, Step Finance, and Drift - and not a single one was a smart contract bug.

Drift's $285M wasn't a program exploit. It was social engineering against the Security Council multisig. Durable nonces kept signatures alive while signers got compromised one by one.

We keep auditing contracts. Attackers keep walking around them.

The unsolved problem in DeFi security isn't code - it's proving that a human actually looked at what they signed.

[Nethny @nethny_dev]

@SolanaFloor @Ledger @DriftProtocol @P3b7_ The scariest part is the consistency. Bybit ($1.5B), Step Finance ($40M), now Drift ($285M) - same playbook every time: skip the contracts, compromise the humans.

$1.8B drained in a little over a year and not a single line of Solidity or Rust was the problem.

[Nethny @nethny_dev]

the pattern across Bybit, Neutrl, Drift is the same: the weakest link isn't the contract, it's human attention under pressure. maturity as an industry probably means hardware that enforces deliberate signing - not just "be more careful." you can't opsec your way out of sophisticated social engineering forever.

[Nethny @nethny_dev]

PDAs solve the tx mechanics. but the Drift attack wasn't a tx problem - it was a human attention problem. signers were compromised before they touched any nonce. what if the hardware required proof you were deliberately focused before authorizing? you can replace the primitive, but you can't patch inattention at the protocol level.

[Nethny @nethny_dev]

pager duty is still reactive - someone has to notice and respond. what if the signing device required hardware-attested proof of deliberate attention before authorizing? attacker compromises the signer, but can't fake the 8-minute focus window. Seed Vault on Seeker could enforce this today. detection becomes irrelevant if signing is physically gated.

[Nethny @nethny_dev]

The real gap isn't multisig - it's that signing proves key access, not deliberate human attention. What if high-value txs required hardware-attested proof that all signers were simultaneously in verified focus for the last 8 minutes? Gaze-tracked, Seed Vault signed. Social eng gets credentials, not temporal proof of collective deliberation. Is anyone building attestation at the biometric layer?

[Nethny @nethny_dev]

@cavemanloverboy [REDACTED] is just "ask again never"
but at least Seeker exists. that's more concrete than most
crypto roadmaps rn 💀

[Nethny @nethny_dev]

This is exactly what on-chain focus verification needs. If
MetaTimer can predict LLM reasoning latency with 6% error,
imagine using that same chain-of-thought decomposition to
validate real focus signals on Seeker. Semantic complexity
→ attention depth. Have you thought about that layer?

[Nethny @nethny_dev]

@joshyote Scale without stickiness is just noise. What % of those 10B
are MEV sandwiches vs actual user behavior? If we solve
that, Solana really is bigger than Earth.