Edward

Cutting through Profits | Research at @HyperliquidX 🔬 | No Paper Hands around here | #NFA

[Edward @Defi_Edward]

You’ve seen hundreds of token launches.

99% fade once the hype dies.

But not @WalletConnect because it was designed to build the next phase of onchain UX.

Here’s why $WCT staking matters 🧵👇🏻

[Edward @Defi_Edward]

So, @StripeCrypto gave agents a stablecoin budget.

The merchant never sees it.

A consumer adds dollars by debit card.

Sstripe converts them into a dollar stablecoin on tempo, held in a privy non-custodial wallet.

When the agent buys, an issued virtual debit card draws on that balance.

The merchant gets an ordinary card payment and needs no stablecoin support.

So, the stablecoin is the funding layer and the card is the acceptance layer.

The per-purchase limit is the control: under it the agent buys, over it the user approves.

Stripe says agent purchases through link grew 38x over the month to late october.

[Stripe Crypto @StripeCrypto]

Stripe Crypto brings together everything needed to enable agentic commerce.

From stablecoins on @tempo to secure non-custodial wallets powered by @privy_io, here's what's powering Link balance for AI agents, starting with @instinct. https://t.co/yn07R9uE1F

[Edward @Defi_Edward]

Grayscale’s bittensor trust can create shares... but it cannot redeem them.

The 10-q, signed november 5: gtao trades on otcqx, the trust is not accepting redemption requests, and its shares have traded at premiums to the tao inside, at times substantial.

Demand for the wrapper kept growing.

Shares outstanding went from 1.88 million in december to 3.34 million on november 2.

The trust returned 47% in q3 on nav.

Creation and redemption are what keep a fund’s price near its nav.

Here only one half exists, and only through a single authorized participant.

[Edward @Defi_Edward]

Deepmind’s agents found two weaknesses in rsa-1024.

None of it touches bitcoin. its keys are not rsa, and rsa-2048 is untouched.

The headline number is 1,600x, and that one is a theoretical estimate.

The measured one is 3.5x less cpu, from two agents and about 20 short messages from a human, in under four days.

Most of us file bitcoin’s cryptographic risk under quantum. kohli’s line is the one i’d keep: don’t assume the hardness of schemes we’ve relied on.

My position: no evidence this reaches ecdsa. I’d still like a bitcoin dev to say so out loud.

[Google DeepMind @GoogleDeepMind]

We’re sharing new research demonstrating how AI agents can accelerate the science of stress-testing encryption.

Guided by humans, our models uncovered two major weaknesses in legacy 1024-bit RSA cryptosystems.

Our VP Science @Pushmeet explains more ↓ https://t.co/SqNkBbkih2

[Edward @Defi_Edward]

Cisco talos found a stealer that asks a language model which victims are worth robbing.

Volta is a windows infostealer: it takes credentials, sessions and wallet seed phrases.

For wallets it goes further and runs an embedded cracker against the vault.

Then it builds a dossier, enriches it with stolen sessions, and posts it to openrouter.

The prompt asks for a target assessment: what they’re worth, ranked, and their opsec.

Default model: claude opus 4.6. if the reply contains refusal phrases like “i can’t”, it falls over to kimi k2, then gemini 2.5 flash.

That chain shows the author expected refusals while talos did not track whether volta is used in the wild.

[Edward @Defi_Edward]

Agent bazaar’s tracker, using coinbase usage counts for october 31, shows 18,057 listed x402 resources.

228 have 20 or more unique paying agents: about 1.3%.

Per one seller’s field notes, the big discovery catalogs list a resource only after its first settled payment.

A new seller can’t settle a payment until someone finds it, and nobody finds it until a payment settles.

Demand clusters in onchain data, rpc and token security checks. The top endpoint has about 480 paying agents at $0.02 a call.

In a market where discovery follows settlement, the sellers who already have buyers are the ones who stay visible.

[Edward @Defi_Edward]

Ai agents did most of the work to formally verify ethereum's fast confirmation rule in lean.

Four to five days, mostly unsupervised.

Then researchers adapted the model to the upcoming fork. the proof would not go through, and the agents showed exactly why.

A fork-choice bug in the spec itself. it is now fixed.

Two limits worth stating. Lean re-checks every proof, so the agents' output does not need to be trusted.

And the proof models the consensus spec the clients are built against, so it speaks for the spec, not for any one client.

[Josh Rudolf | Ethlabs @joshrudolf]

https://t.co/JTZOGwesr2

[Edward @Defi_Edward]

The cftc held its first forum on ai agents wednesday. chairman selig raised three questions, and the answers are still open.

1. Does an agent's developer have to register?

2. Who is liable when the software goes off course?

3. How does a regulator know on whose behalf an agent, or a group of agents, is trading?

His example was a refrigerator staking ether.

He also flagged that agents on similar models can end up trading the same way at the same time.

The third question matters most on a public chain.

An agent shows up as a wallet address, and nothing in an address says whose it is

[Edward @Defi_Edward]

$USDT has an issuer who can freeze it, $ETH doesn't.

Per llamarisk, a wallet in the binance exploiter's cluster took an aave loan, borrowing eth against usdt collateral while tether and circle were freezing funds.

Borrowing against a freezable asset turns it into an unfreezable one without selling anything.

Exactly the move musk was pointing at when he said superintelligent agents would do this someday.

Reports are clear no ai was involved here.

The mechanism is an ordinary lending move, no ai required.

[Elon Musk @elonmusk]

@cryptoblaney Super intelligence agents will at some point do this to gain currency for themselves, not a human

[Edward @Defi_Edward]

@0xJigen 100% true, great projects g

You should check $PAPA & $ANTI

[Edward @Defi_Edward]

Jpmorgan, bank of america, and citigroup are restricting third-party ai agents from touching customer accounts this week.

Reportedly because the agents log in with legitimate credentials but act without clear human approval.

@binance founder changpeng zhao responded directly: crypto becomes the preferred payment rail for ai agents precisely because banks can't adapt fast enough.

Worth separating the actual claim from the opportunism here.

Banks aren't blocking agents because crypto solved a problem they couldn't, they're blocking agents because nobody's built a permission system either side trusts yet.

Crypto hasn't solved that either, it's just not pretending the old rails were built for this in the first place.

[Edward @Defi_Edward]

zkML proofs are now settling real payouts on prediction markets.

A short cryptographic proof confirms a specific model produced a specific output for a specific input, and the market pays out based on that proof instead of trusting the operator's word.

Worth being precise about what that actually removes.

It removes the operator's ability to lie about what the model said.

It doesn't remove the possibility that the model was simply wrong.

A confidently incorrect output gets the exact same cryptographic guarantee as a correct one, the proof only ever covers the computation, never the truth of what came out of it.

[Edward @Defi_Edward]

Grayscale's head of research published a report yesterday.

It named three ai infrastructure protocols worth watching beyond the usual bittensor, near, and world. one of them, pearl, is already up 580% over the past three months, per coingecko data cited in grayscale's own piece.

Worth noting what that actually tells you and what it doesn't. a name showing up in an institutional research report after a 580% run isn't the same as catching it before the move.

The thesis might still be right going forward, the entry point just already moved a long way without you.

[Edward @Defi_Edward]

@SherifDefi @ethena Goated

[Sherif @SherifDefi]

Thanks @surgence_io 🫱🏻‍🫲🏼

[Edward @Defi_Edward]

@SherifDefi @surgence_io Alphaaa

[Edward @Defi_Edward]

A researcher scanned the internet for exposed ai agent control panels.

Found about two dozen instances of cairn, an agent framework attackers like, with no login required.

Over 1,200 sites got breached this way. Average break-in time: under 23 hours.

These panels commonly answer on port 8000.

Check whether your agent framework's panel is actually reachable from the open internet.

A password screen that isn't there does nothing.

An exposed panel with no auth check is the same as no security at all.

[Edward @Defi_Edward]

Honestly, @thdxr drew a hard line here: the console is a separate product from the actual app.

And he's right about that distinction.

It doesn't change who's actually on the hook though.

A user who pasted their openai or anthropic key into that console got hit regardless of which product technically failed.

"Those keys never come to us" protects opencode's liability position.

It does nothing for the person whose account just ran up unauthorized charges from a service they trusted with a credential.

[dax @thdxr]

@lydiahallie note this is the OpenCode Console, not OpenCode itself, those keys never come to us

[Edward @Defi_Edward]

@0xJigen @tryantimatter 👀👀

[Edward @Defi_Edward]

68.9% of crypto traders use ai for research.

Only 13% have actually tried letting an agent execute trades.

Meanwhile ai agent token prices fell 7% in 24 hours even as trading volume in that category rose 28.4%.

It’s a structural mismatch worth naming precisely.

Rising volume with falling price usually means conviction weakening while activity increases, exactly what you'd expect from a sector people are trading more but trusting less.

The market for "ai handles your trades" tokens is pricing in doubt about the thing it's named after.

[Edward @Defi_Edward]

aave v4 confirmed this week that it relies exclusively on chainlink for all pricing.

It includes also the new equities and real-world asset products.

Formally verifying the lending and liquidation logic proves that logic is internally consistent, that collateral ratios get calculated correctly given whatever price comes in.

It proves nothing about whether that price is actually correct. a liquidation engine can be mathematically flawless and still liquidate the wrong position at the wrong time if the single oracle feeding it reports a bad number.