BTC $64,014.42, ETH $1,872.54, SOL $73.97, XRP $1.076, HYPE $55.45, DOGE $0.070378, AAVE $90.17, ZEC $499.17.
Wednesday’s useful signal is operating leverage.
The August 3-4 digests already covered CLARITY procedure, Coldcard loss mechanics and exchange-flow behavior, Spark’s backend pivot, Fun’s funding abstraction, traditional-asset perps, Robinhood’s U.K. registration, Ripple tokenization plumbing, Bithumb’s IPO path, prediction-market volume, Rails and Chrome security, Circle’s downgrade, BlackRock reserve products, Strategy’s BTC sale, ZeroStack’s going-concern warning, American Bitcoin’s power story, bitcoin basis compression, yen intervention, N-able, Solana fee burns, FXRP collateral, Hashdex’s ETF closure, Boltz, Telegram’s app-store gate, and the last four GitHub repo sets.
This morning moves away from that mix. The new board is about who controls the interfaces beneath trading, settlement, compute, security intake, and agent work.
Wells Fargo joined JPMorgan and Citi in tokenized settlement rails. Polymarket reportedly returned to investors at a $20 billion valuation. Robinhood’s event-contract business out-earned its crypto business in Q2. Dinari opened tokenized S&P 500 stocks to eligible U.S. self-custody wallets using USDC. BitGo moved $7.4 billion of WBTC from LayerZero’s OFT standard to Chainlink CCIP. Texas ordered an audit of data-center interconnection requests, turning approved megawatts into scarcer assets. Hut 8 missed revenue estimates while growing its AI data-center pipeline to 8.7 GW. Apple capped vulnerability submissions after AI-generated bug reports flooded its security queue. Five men were convicted in London after two crypto-rich victims were held for 52 hours. GitHub’s fresh board points to bounded software factories, long-horizon computer-use harnesses, and MCP-based desktop control.
That is a better Wednesday mix. Less “which wrapper survived yesterday?” More “which operating layer gets paid when crypto and AI stop being side projects?”
Price snapshot via Coinbase spot prices for BTC and ETH, plus CoinGecko simple-price data for SOL, XRP, HYPE, DOGE, AAVE, and ZEC around 05:58 HKT.
1. Wells Fargo Joined The Settlement-Rail Race
Benzinga reported that Wells Fargo has joined JPMorgan and Citi in the race to tokenize Wall Street settlement rails.
That is the bank story worth watching because it isn’t a crypto listing or a balance-sheet trade.
The pitch is payments running on the bank’s own blockchain and routing through the client interface institutions already use. That sounds dull until you remember how much of finance is still a fight over who owns settlement timing, reconciliation, collateral movement, cash legs, and operational data.
JPMorgan’s Kinexys and Citi’s tokenized-deposit work already pushed large banks into private settlement networks. Wells Fargo entering the same lane makes this less like an innovation-lab experiment and more like a competitive requirement.
The public-chain question remains open. Banks may tokenize internal settlement first, then connect outward only when legal, liquidity, and customer-service risk is contained. That is still meaningful for crypto because the endpoint changes. If bank clients can move tokenized cash inside familiar portals, crypto-native rails have to compete against default workflow, not just slower legacy systems.
Tokenization keeps sounding like an asset story. The bigger prize is who owns the operating rail.
2. Polymarket Is Pricing Itself Like Market Infrastructure
CoinDesk reported that Polymarket is seeking fresh capital at a reported $20 billion valuation.
The timing is aggressive.
CoinDesk says Polymarket closed a round at a $15 billion valuation in April, including a $600 million investment from Intercontinental Exchange, the NYSE owner. A few months later, the target is higher again. The explanation is that prediction markets are no longer being valued as a crypto novelty.
CEO Shayne Coplan keeps framing Polymarket as an information market rather than a betting site. That matters because Coinbase, Robinhood, and Kalshi are now attacking the same category from different regulatory and distribution angles.
The key split is venue identity. Kalshi operates a federally regulated U.S. exchange. Polymarket uses blockchain settlement and crypto-native liquidity. Robinhood already has consumer trading distribution. Coinbase has crypto accounts and ambitions to broaden its trading menu.
The market is asking a sharper question: will prediction markets become a feature inside brokers, a standalone exchange category, or a crypto settlement layer with better event design?
A $20 billion valuation assumes the answer is bigger than election betting.
3. Robinhood Replaced Crypto Beta With Event Contracts
The Block reported that Robinhood and Coinbase both saw weaker crypto trading activity, but Robinhood offset the hit with options, equities, and prediction markets.
The numbers are blunt.
Robinhood’s crypto transaction revenue fell to $100 million in Q2, down 37.5% year over year. Crypto is now only 13% of Robinhood transaction revenue after volume fell another 23% quarter over quarter.
The company still matched its Q4 2025 transaction-revenue record at $776 million. Options revenue hit $342 million. Equities hit $129 million. Event contracts generated $156 million on 13.6 billion contracts traded, which means prediction markets out-earned crypto for the first time.
Coinbase had the opposite problem. Q2 revenue fell to $1.22 billion, its weakest since Q3 2024. Trading volume fell 24% quarter over quarter, and transaction gross margin compressed to 68% from 74%.
This is a cleaner distribution signal than another product launch.
Robinhood built a way to monetize user speculation when crypto slowed. Coinbase still looks more tied to the crypto trading cycle. The winning consumer-finance app may be the one that can route attention across options, equities, sports, politics, rates, and crypto without depending on one market’s volatility.
4. Dinari Brought Tokenized Stocks To U.S. Self-Custody
The Block reported that Dinari launched tokenized S&P 500 stock trading for eligible U.S. investors using USDC in self-custody wallets.
That is a different tokenized-equity story from the offshore wrappers that dominated earlier cycles.
Dinari says users can buy and sell more than 700 tokenized stocks, including every S&P 500 name, through a Circle partnership. Its dShares are backed by corresponding securities held in qualified custody.
The caveat is just as important as the launch. Dinari warned that tokenized-security markets may be limited, which can make selling harder at the desired time or price. It also said the legal treatment of tokenized securities and digital assets is still developing.
That is the real product boundary.
Tokenized stocks are useful only if the user can understand the claim: who holds the underlying security, what rights pass through, how liquidity works, what happens during corporate actions, and whether the wrapper survives regulatory changes.
The launch pushes self-custody closer to U.S. equity exposure. It also drags self-custody into securities-market plumbing that can’t be hand-waved away.
5. BitGo Moved WBTC Into Chainlink’s Security Model
The Block reported that BitGo is moving $7.4 billion of WBTC from LayerZero’s OFT standard to Chainlink CCIP.
That makes WBTC the biggest asset yet in the LayerZero-to-Chainlink migration wave.
The switch follows the $292 million KelpDAO bridge exploit earlier this year. The Block says announced value moving from LayerZero to Chainlink is now near $15 billion. BitGo also plans to use CCIP as the exclusive inter-blockchain standard for future BitGo-issued assets.
The useful detail is the security contrast. Chainlink describes CCIP lanes as secured by at least 16 independent, security-reviewed node operators, while LayerZero’s older configurable DVN model let some clients run weaker setups. KelpDAO reportedly used a 1-of-1 verifier configuration at the time of its exploit.
WBTC still carries its own history: BitGo’s multi-jurisdiction custody arrangement, BiT Global’s key role, Coinbase delisting, and cbBTC competition. The CCIP move doesn’t erase those issues.
It does show where the market is going after large bridge losses. Default security posture matters. Configurability is attractive until users learn the configuration was the product risk.
6. Texas Turned Approved Power Into A Scarcity Asset
The Block reported that Bernstein expects Texas’ state-ordered data-center audit to raise the value of already approved bitcoin mining and AI sites.
Texas Governor Greg Abbott directed the Public Utility Commission of Texas and ERCOT to audit all data-center projects in the state’s interconnection queue. ERCOT then paused its Batch Zero review process. Data-center requests reportedly make up about 90% of ERCOT’s 474 GW interconnect queue.
This is where mining, AI, and politics meet.
Speculative power requests are easy to announce and hard to serve. If Texas slows the queue, developers with pending projects may get delayed. Operators with approved capacity suddenly own something harder to replace.
Bernstein called out Cipher, CleanSpark, and Core Scientific as more exposed because near-term expansion depends on Texas grid connectivity. It said Terawulf benefits from geographic diversification, while Riot and IREN benefit from already approved ERCOT footprints.
The crypto read is simple: miners are no longer valued only on hashrate and bitcoin holdings. Approved megawatts, grid status, local politics, and interconnect credibility now drive the trade.
AI made power scarce. Bitcoin miners already sitting on real capacity may get paid for being early.
7. Hut 8’s AI Pipeline Grew While Earnings Missed
The Block reported that Hut 8 shares fell after a Q2 revenue miss, even as the company’s AI data-center pipeline expanded.
Revenue was $74.9 million, up from $41.3 million a year earlier but below roughly $80 million expected by Wall Street. The company posted a $177.1 million net loss, including $138.6 million of unrealized digital-asset losses.
The investor disappointment is real. The strategy shift is too.
Hut 8 said its development pipeline reached about 8.7 GW, up roughly 300 MW quarter over quarter. It is evaluating 11 sites under exclusivity, each averaging more than 650 MW. Management also reiterated its 1 GW Beacon Point AI campus commercialization and a 949 MW contracted-capacity portfolio with about $26.6 billion in expected contract value.
CEO Asher Genoot said future bitcoin exposure should primarily sit in American Bitcoin, Hut’s majority-owned subsidiary. He described bitcoin on Hut 8’s own balance sheet as another asset, closer to cash than company identity.
That is the quiet change.
Some miners are becoming AI-infrastructure companies with bitcoin exposure. Others are bitcoin companies with power assets. The market is still deciding which version deserves the higher multiple.
8. Apple’s Bug Queue Hit The AI Flood
Decrypt reported that Apple capped how many vulnerability reports a researcher can keep open after a surge of AI-generated submissions.
The cap caught a real report in the net.
Milan security startup Bynario told the Financial Times it used ChatGPT to find more than 50 macOS bugs in three weeks, including a privilege-escalation chain that could give an attacker full control of a Mac. The company said it couldn’t file the report because Apple’s portal had refused further submissions. Bynario’s CEO estimated the exploit’s criminal-market value at $100,000 to $200,000.
Apple said it is now in contact with the firm and reviewing the work. It also said researchers can request higher limits.
This is the AI-security bottleneck in one story.
AI can find real bugs. It can also generate fake reports at scale. Maintainers then need human time to separate signal from junk. Bugcrowd said submissions through its platform more than quadrupled in three weeks in March, and most were fake.
The hard part isn’t only model capability. It’s triage capacity. Security programs now need intake systems that assume both attackers and researchers can produce more volume than humans can read.
9. London Showed The Physical Crypto Threat Model
Decrypt reported that five men were convicted of false imprisonment after two French crypto millionaires were held for 52 hours in an East London flat.
The victims were ambushed in Shadwell and taken to Canning Town. Prosecutors said their lifestyles, much of it posted on social media, may have attracted the gang’s attention. The attackers demanded $150,000 and received $30,000.
Police traced the flat after a friend left a phone in the car. Two men were convicted of conspiracy to blackmail and false imprisonment. Three others were convicted of false imprisonment or admitted it. The man police say directed the plot from abroad has not been arrested.
This belongs in a market digest because the industry still over-indexes on software threat models.
Hardware wallets, multisig, passkeys, withdrawal limits, and transaction monitoring all matter. They don’t solve public wealth signaling, travel exposure, social-media leakage, home-address leakage, conference routines, and coercion.
Crypto made self-custody possible. It also made some people walking withdrawal endpoints.
Operational security has to include the body, not just the browser.
10. GitHub Trending - Bounded Agents, Long-Horizon Harnesses, And Desktop MCP
The featured-repo tracker ruled out the August 3-4 sets, including yc-software/qm, firecrawl/pdf-inspector, QwenAudio/qwen-audio-agent, microsoft/skill-recorder, sqliteai/waste, Kritt-ai/open-kritt, TencentCloud/Octop, HKUDS/Vibe-Trading, marianfoo/sap-ai-mcp-servers, antirez/ds4, livekit/agents, and microsoft/generative-ai-for-beginners.
Fresh picks from GitHub search and repo metadata:
disler/super-simple-software-factory has 321 stars and was created on August 2. It packages repeatable agents-plus-code workflows as a skill that can be stamped into repos. The interesting part is the constraint: deterministic Python owns the graph, while coding agents run as bounded nodes. That is exactly where serious agent workflows are going. Let the agent reason, but don’t let it own the whole process.
AMAP-ML/LongHorizon-Harness has 154 stars and was created on August 4. It is a computer-use harness for agents working across desktop apps and CLIs over extended periods. The signal is durable state: long tasks need checkpoints, independent auditing, recoverable progress, and fresh-context execution. Chat UX isn’t enough when the job spans hours.
mrpulor-gh/nuphus-mcp has 100 stars and was created on August 1. It is an MCP server for desktop automation: screen, windows, mouse, keyboard, and Chrome control over stdio. That sits in a crowded but important lane. Agents are moving from text tools into operating-system control, and the MCP layer is becoming the permission boundary.
Morning Read
Read Wells Fargo’s settlement-rail move, then read Polymarket’s reported $20 billion valuation target, then read Robinhood’s event-contract revenue split.
The number to remember is $156 million.
That is Robinhood’s Q2 event-contract revenue, enough for prediction markets to out-earn crypto inside the app. The second number is 474 GW, because Texas’ interconnection queue shows how absurd the AI-and-mining power race has become.
Wednesday’s read is operating leverage. Tokenization, prediction markets, wrapped bitcoin, AI data centers, bug bounty queues, physical security, and desktop agents all point in the same direction: the product headline matters less than the control surface beneath it.
Who owns the rail? Who owns the queue? Who owns the permission boundary? That is where the next margin fight sits.
Evening Update - 18:15 HKT
BTC $64,050.21, ETH $1,869.01, SOL $73.94, XRP $1.061, HYPE $57.09, DOGE $0.069699, AAVE $91.44, ZEC $516.18.
Wednesday evening moved from operating leverage to distribution friction.
The novelty gate ruled out another broad pass through this morning’s Wells Fargo settlement-rail story, Polymarket valuation chase, Robinhood event-contract split, Dinari tokenized stocks, WBTC’s CCIP migration, Texas power queue, Hut 8’s AI pipeline, Apple’s bug-report flood, London physical-security case, and the morning GitHub trio. It also kept the August 3-4 loop out unless the facts changed the consequence: Coldcard mechanics, CLARITY procedure, stablecoin reserve wrappers, Strategy’s capital stack, tokenized-stock concentration, bitcoin basis compression, app-store gatekeeping, and repeated miner-power framing all needed a better reason to return.
The evening stories clear that bar. Crypto failed to rally even as global equities printed records and oil softened on a possible Hormuz deal. Ethereum researchers proposed burning validator issuance as staked ETH climbs toward roughly half the supply. Samsung wants stablecoin features inside 800 million new Galaxy phones. SpaceX’s first public quarter beat revenue expectations, but its bitcoin stash was marked $540 million lower while AI capex ran hot. Warren and Blumenthal asked the SEC to investigate TRUMP memecoin losses as CLARITY’s ethics fight tightened. Fairshake-backed candidates advanced in Michigan and Washington, turning earlier ad spend into election results. Eliza Labs’ founder declared the token dead after a class-action settlement while keeping the open-source agent framework alive. GLM-5.2 showed how close open-weight models are getting to frontier cyber and bio capability without the same safety layer. ChainDrop compromised more than 1,300 npm packages with valid provenance. GitHub’s fresh board points to mobile QA agents, persistent local workspaces, and prompt-history audits.
That is a sharper evening mix. Less “who owns the interface?” More “what happens when the interface reaches users, voters, validators, phones, CI runners, and public markets?”
Price snapshot via Coinbase spot prices for BTC and ETH, plus CoinGecko simple-price data for SOL, XRP, HYPE, DOGE, AAVE, and ZEC around 18:15 HKT.
11. Crypto Missed The Equity Rally
CoinDesk reported that bitcoin held just above $64,000 while global equities pushed toward fresh records.
That is the market story because the macro tape should have helped.
MSCI’s world index was near another record close. Asia-Pacific equities gained 2.2%. The S&P 500 and Dow had both closed at all-time highs on Tuesday. Brent crude fell toward $78.50 after reports that Washington, Tehran, and Oman were close to a Strait of Hormuz agreement. Treasuries and gold advanced as traders cut rate-hike bets.
Crypto mostly sat there.
Bitcoin was roughly flat on the week and still about 49% below its October peak. Ether was the only major token down on the week in CoinDesk’s snapshot. HYPE was the exception, up around 3% on the day and week.
That points the drag inward. If risk assets rally, oil cools, and rate fear eases while bitcoin doesn’t move, the market is probably processing crypto-specific supply, custody, treasury, and liquidity issues.
The next clean test is whether a confirmed Hormuz deal changes the tape. If it doesn’t, the buyer problem is inside crypto.
12. Ethereum Put Staking Economics On The Knife Edge
CoinDesk reported that six Ethereum researchers proposed gradually burning validator rewards as the staking ratio rises.
The proposal, EIP-8361, would burn a larger share of newly issued ETH as more ETH gets staked. At about 60.25 million ETH, roughly half the supply, the burn would reach 100% of consensus-layer issuance.
That isn’t a small tuning change.
About 41 million ETH is already staked, close to 34% of supply. Another 2.5 million ETH sits in the activation queue, with a wait of six weeks or more. One proposal author projects more than 70 million ETH staked by January 2028 if nothing changes.
The case for the change is decentralization. If staking always pays, more ETH keeps flowing into validators, exchanges, liquid staking providers, and large pools. Past a point, more stake can make the system less healthy by pushing smaller validators out.
The DeFi objection is just as obvious. Borrowed ETH, liquid staking loops, restaking collateral, and ETH carry trades all depend on staking yield being real enough to support leverage.
Ethereum’s next fight isn’t only scaling. It’s also deciding how much monetary policy should punish concentration before DeFi breaks its own assumptions.
13. Samsung Turned Stablecoin Distribution Into A Phone-OEM Fight
CoinDesk reported that Samsung plans to add native stablecoin features to 800 million new Galaxy phones through Samsung Wallet.
That is a different stablecoin story from issuer competition.
Samsung already has crypto wallet infrastructure through Knox and says Samsung Wallet has nearly 19 million users in South Korea. Its wallet is available in 61 countries. The new plan adds fiat-pegged savings and payment accounts directly into Galaxy hardware.
Distribution is the scarce asset. A stablecoin inside a phone wallet doesn’t need the user to download a crypto app, pick a chain, or trust a new exchange account on day one. It starts closer to cards, payments, rewards, checkout, and device identity.
The infrastructure half matters too. Samsung affiliates agreed in May to buy a 4% stake in Dunamu, the Upbit operator, for $408 million. Samsung SDS has described the investment as a digital-asset infrastructure move, including stablecoins and AI-powered payments.
Apple and Google still own the bigger mobile operating-system gates. Samsung is trying to own a default financial surface inside its own hardware.
That is where stablecoin adoption could get less crypto-native fast.
14. SpaceX Mixed Bitcoin Marks With AI Capex
CoinDesk reported that SpaceX posted $7.8 billion of Q2 revenue in its first public quarterly report, beating Wall Street’s $6.9 billion expectation.
The headline was good. The capital story was messier.
SpaceX narrowed its net loss to $541 million from $1.0 billion a year earlier. Adjusted EBITDA nearly tripled to $3.5 billion as launch, Starlink, and AI businesses grew. The company still held 18,712 BTC, but its digital-asset value fell to $1.10 billion at June 30 from $1.64 billion at the end of 2025.
The bitcoin mark was not the only reason investors flinched. SpaceX spent $18.4 billion on capital expenditures during the quarter, above the $13 billion analysts expected, as it kept pouring money into AI infrastructure.
The stock fell after hours despite the revenue beat. An insider-share unlock also arrives on August 6, when roughly 912 million employee and early-backer shares become eligible for sale.
This is what bitcoin treasury exposure looks like inside a serious public operating company. The BTC mark is one line. AI capex, float expansion, EBITDA, and public-market expectations decide whether investors care.
15. TRUMP Made CLARITY’s Ethics Problem Harder To Ignore
The Block reported that Senators Elizabeth Warren and Richard Blumenthal asked the SEC to investigate President Donald Trump’s TRUMP memecoin.
This is the evening’s clean policy pressure point.
The senators cited reports that nearly 1 million wallets have lost money since the token launched in January 2025, with combined losses around $3.81 billion. They also said Trump made $636 million from the memecoin and asked the SEC to determine whether securities laws were violated.
The request may not lead to fast SEC action. The agency has already said many memecoins sit outside its securities remit.
The timing still matters. CLARITY negotiators are trying to finish ethics language before senators leave Friday for August recess. A prior compromise would have limited senior officials and spouses from issuing or sponsoring digital assets, but Democrats objected because the structure left practical gaps.
The token is no longer only a campaign-finance optics problem. It’s also directly entangled with whether enough Democrats can vote for market-structure legislation.
Crypto wants rules. The president’s own coin keeps making the rulemaking harder.
16. Fairshake Converted Ad Spend Into Primary Wins
The Block reported that several Fairshake-backed candidates advanced in Michigan and Washington primaries on Tuesday.
That is the changed fact after Sunday’s Michigan PAC-spending story.
In Michigan, Republican Rep. Bill Huizenga won his party’s nomination after Defend American Jobs, Fairshake’s Republican-aligned PAC, spent nearly $512,000 supporting him. In Washington, Democratic Reps. Suzan DelBene, Kim Schrier, and Marilyn Strickland won their primaries, while Republican Amanda McKinney also advanced.
Fairshake affiliates backed all four Washington candidates. Protect Progress spent $113,120 supporting DelBene, $105,040 supporting Schrier, and $103,020 supporting Strickland. Defend American Jobs spent $506,917 backing McKinney.
The dollar amounts are not huge next to national races. That is exactly why they matter.
Primary elections are where concentrated industry money can move the most. Fairshake has a reported $193 million midterm war chest and is using it to make crypto votes less abstract for members of Congress.
The risk is blowback. Every win also gives critics a cleaner attack line: the industry isn’t only lobbying the rulebook, it’s also buying insurance around the people who write it.
17. Eliza Killed The Token And Kept The Agent Framework
The Block reported that Eliza Labs founder Shaw Walters declared the project’s associated token dead and said the related foundation is winding down.
That is an AI-agent story wearing a DeFi wound.
ElizaOS remains an open-source framework for building and managing AI agents. The token began as ai16z on Solana in October 2024, with a pitch around an AI-managed DAO and venture-style fund. After a rebrand and migration, holders sued.
Burwick Law filed a federal class action in April, alleging false advertising, deceptive practices, negligent misrepresentation, and unjust enrichment. The complaint said the project marketed itself as autonomous and AI-managed while insiders allegedly controlled key decisions. Walters said the team settled by transferring the remaining treasury and available funds.
He also said he would not support or launch another token tied to Eliza.
That split is useful. The software can live. The token layer can die. In fact, the token may have become a liability for the work.
Open-source agent frameworks don’t automatically need a coin. Sometimes the coin only adds governance theater, holder expectations, legal exposure, and community debt.
18. Open-Weight AI Is Closing The Capability Gap Before The Safety Gap
TechCrunch reported that Z.ai’s GLM-5.2 is only a few months behind OpenAI and Anthropic frontier systems on cyber and bio capabilities, according to a SaferAI report.
The uncomfortable part is refusal behavior.
SaferAI tested GLM-5.2 through Z.ai’s public API and said it refused none of the offensive cyber or dual-use biology tasks it received. By contrast, Claude Opus 4.7 refused so consistently that SaferAI could not complete CyberGym on it.
That doesn’t mean open-weight models are bad. It means the policy problem is different.
Closed model providers can lean on classifiers, refusal training, hosted API controls, usage monitoring, and access tiers. Those controls are imperfect, but they exist at the service layer. Once weights are public and local, a user can remove system prompts, fine-tune behavior, run without monitoring, and ignore hosted guardrails.
The same capability helps defenders too. Hugging Face argued that strong open models can help identify and fix vulnerabilities at scale.
Both claims can be true. Open weights democratize defensive power and offensive reach at the same time.
The release question is no longer “can open models compete?” It is “what safety layer survives after download?“
19. ChainDrop Made Provenance Feel Less Comforting
BleepingComputer reported that ChainDrop compromised more than 1,300 npm packages with a combined 2 billion monthly downloads.
This is the developer-security story that should bother every crypto and AI team.
The attack began after the GitHub account of Keyv’s maintainer was compromised. It then spread through packages tied to Keyv, Cacheable, flat-cache, file-entry-cache, and organizations including Deliveroo, Picsart, Qlik, and ServiceTitan. Aikido said at least 868 packages across 1,381 versions were compromised.
The packages were built and published through legitimate GitHub Actions workflows, so the malicious releases carried valid provenance information. That is the nasty detail.
The payload ran on npm install, downloaded the Bun runtime, and executed an infostealer that searched developer machines and CI runners for environment variables, GitHub tokens, npm tokens, cloud secrets, Kubernetes secrets, Vault tokens, database credentials, private keys, Stripe, Slack, Twilio, Azure, and GCP credentials.
If an affected version landed in a build, uninstalling the package isn’t enough. The machine or runner has to be treated as compromised and secrets need rotation.
Provenance tells you where a package came from. It doesn’t prove the maintainer’s account, workflow, or release path was clean.
20. GitHub Trending - Mobile QA Agents, Persistent Workspaces, And Prompt Audits
The featured-repo tracker ruled out the August 3-5 sets, including yc-software/qm, firecrawl/pdf-inspector, QwenAudio/qwen-audio-agent, microsoft/skill-recorder, sqliteai/waste, Kritt-ai/open-kritt, TencentCloud/Octop, HKUDS/Vibe-Trading, marianfoo/sap-ai-mcp-servers, antirez/ds4, livekit/agents, microsoft/generative-ai-for-beginners, disler/super-simple-software-factory, AMAP-ML/LongHorizon-Harness, and mrpulor-gh/nuphus-mcp.
Fresh picks from GitHub search, GitHub Trending, and repo metadata:
callstack/agent-device has about 4.0K stars. It gives coding agents a CLI for inspecting, controlling, and verifying iOS, Android, TV, web, macOS, and Linux apps, with evidence capture for review. That is the next testing gap: agents can write app code faster than humans can poke every mobile state, so verification needs to move closer to the device.
kirodotdev/KiroCrew has 735 stars. It packages a persistent development workspace with local or remote execution, memory, recurring jobs, heartbeats, skills, MCP, apps, approvals, sandboxing, sensitive-path controls, and audit logs. The interesting part is the operating model. Useful agents need state, schedules, permissions, and review surfaces that outlive a single chat.
WEIFENG2333/phistory has 477 stars. It archives prompt snapshots from coding-agent CLIs including Claude Code, Codex, Antigravity, Grok Build, Kimi Code, OpenClaw, Hermes, opencode, and others. That sounds niche until you realize prompts are now part of runtime behavior. Versioned prompt history is becoming audit evidence.
Evening Read
Read Ethereum’s EIP-8361 staking proposal, then read Samsung’s stablecoin distribution play, then read ChainDrop’s npm supply-chain attack.
The number to remember is 800 million.
That is the number of new Galaxy phones Samsung wants to ship with stablecoin features. The second number is 1,300, because ChainDrop shows how quickly a compromised maintainer path can turn into a package-registry problem with valid-looking provenance.
Wednesday evening is about distribution friction. Crypto wants more default surfaces: phone wallets, public equities, agents, package managers, voters, validators, and CI runners. Every default surface adds reach. Every default surface also adds a new control problem.