Tag: Business – Decrypt

Coinkite believes an attacker used AI to find a flaw that has cost owners of its Coldcard hardware wallets tens of millions of dollars in Bitcoin, and says its own AI review of the same code weeks earlier turned up nothing. 

The hardware wallet manufacturer published an advisory for its Mk3 and a technical breakdown on Thursday, after learning that seeds generated by its devices were far more guessable than intended.

The losses to the flaw, which was exploited early Friday, are estimated at 594 BTC, around $38 million. Funds were drained from roughly 500 wallets inside 25 minutes, with 562 BTC since consolidated into a single address.

Coinkite said it has to assume “someone used AI to review previous versions of our firmware” in order to uncover the flaw. The firm said it had run one of the best available models over its own code a few weeks earlier, and the model “did not find this bug or anything serious.” Attackers and defenders have the same tools, it wrote, but this time “it did not help us, and only helped the bad guys.”

What went wrong

Coldcard’s firmware calls a function to fetch randomness, and two implementations of it sat in the codebase with identical signatures: the hardware generator Coinkite wrote, and a software fallback inherited from MicroPython. A preprocessor guard checked only whether a setting was defined, without testing its value, so the build completed against the fallback without complaint. Seed generation had been drawing on it since a March 2021 migration.

Every current model is affected to some degree. Coinkite estimates the effective search space for an Mk3 seed at about 40 bits, against the 128 a seed is meant to have. Extra entropy from the secure elements on the Mk4, Q and Mk5 lifts theirs to roughly 72 bits, which the company says materially improves the position without reaching the target. Tapsigner, Opendime and Satscard use different code and are unaffected.

What owners must do

Coinkite has shipped an emergency hotfix, version 5.6.0 for the Mk4 and Mk5 and 1.5.0Q for the Q. Updating does not repair a seed already created on affected firmware. Owners need a new seed generated on patched hardware, and the company recommends a strong BIP-39 passphrase, at least 99 dice rolls, or both. Mk3 owners, whose model is out of support, are pointed to a separate migration path.

A seed created on an affected Coldcard stays weak after being restored to another brand’s device, a point rival hardware wallet manufacturer Trezor made while telling its own users their funds are safe. Block, which published an independent analysis on Friday, said none of its products are affected, and its hardware lead Max Guise urged anyone exposed to move funds as soon as they safely can.

Coinbase reported $1.22 billion in second-quarter revenue on Thursday, down 14% from the previous quarter, and a net loss of $359 million.

Coinbase shares fell sharply in after-hours trading, losing roughly 5% of its stock price, after missing on earnings estimates for the quarter. The company was expected to bring in $1.29 billion in Q2.

According to Coinbase, total crypto spot trading volume declined more than 20% from the previous quarter as crypto asset prices fell and market volatility reached multi-year lows. Transaction revenue totaled $599 million—lower than the expected $628 million.

Subscription and services revenue totaled $555 million, representing 48% of net revenue. Coinbase said the figure was below its previously forecast range of $565 million to $645 million because certain USDC-related commercial agreements closed later than expected and lower crypto asset prices reduced staking revenue.

Stablecoin revenue totaled $292 million. Average USDC held across Coinbase products reached a record $20 billion during the quarter, representing more than 30% of USDC in circulation at quarter-end. Coinbase also said 88% of net revenue came from sources other than Bitcoin spot trading, compared with 45% in the second quarter of 2020.

Coinbase said its crypto trading market share reached a record 10.3% during the quarter, its third consecutive quarter of market share gains. The company said it gained share in both spot and derivatives trading.

In a bright spot for the company, prediction markets contracts and revenue grew 106% from the previous quarter and exceeded a $100 million quarterly annualized net revenue run rate, according to Coinbase. Average Borrow/Lend balances increased by more than $1 billion from a year earlier to $1.49 billion. The company also said the conditions for its commercial agreement with Circle to renew in August automatically had been met.

The earnings report follows a busy second quarter for Coinbase.

In May, the company became the first U.S. crypto exchange cleared to offer customers access to offshore crypto perpetual futures through its Deribit subsidiary. In June, Coinbase launched Coinbase for Agents, a platform that lets AI agents trade crypto, make payments and manage portfolios on users’ behalf. Later that month, the company announced plans to launch tokenized stock trading, crypto and equities options, along with new lending and rewards products.

Coinbase ended the quarter with $8.6 billion in cash and cash equivalents and $10 billion in total available resources. During the quarter, the company repurchased 814,000 Class A shares. Year to date, it has repurchased nearly 7 million shares for $1.2 billion, leaving about $2 billion remaining under its share repurchase authorization.

For the third quarter, Coinbase said transaction revenue totaled approximately $130 million through July 26. The company expects subscription and services revenue between $500 million and $580 million and adjusted expenses between $980 million and $1.08 billion.

Despite the weaker-than-expected earnings report, Coinbase CEO Brian Armstrong remained optimistic about the future of the company.

“Coinbase is no longer a bet just on the price of Bitcoin,” he said during the earnings presentation. “All of financial services are getting updated by crypto technology, whether that’s trading or payments or lending. And Coinbase is the best-positioned company in the world to power this. And of course, this next frontier is going to be agentic finance, where we’re an early leader.”

Some Windows 11 users are discovering a new Microsoft app they never installed. Called OneDrive Photos, it automatically indexes local images and offers AI-powered search and optional facial grouping.

The rollout, first spotted by industry publication Windows Latest, comes via Windows Update or updates to the existing OneDrive client rather than as a separate download.

According to Windows Latest, OneDrive Photos can detect and display images stored locally on a PC even if the user is not signed into a Microsoft account. After signing in, the app enables AI-powered search that can locate photos using natural-language queries and optical character recognition.

The app also includes an optional People feature that groups similar faces, which are visible only to the user, aren’t shared with third parties, and can be deleted by disabling the feature. However, Windows Latest noted that the feature requires users to opt in because facial data may be considered biometric information in some jurisdictions.

The news comes two years after Microsoft was forced to overhaul its AI-powered Recall feature following a privacy backlash that prompted the company to make the screenshotting tool opt in by default.

More recently, in May, Google began downloading its roughly 4GB Gemini Nano AI model to eligible Chrome installations through browser updates to power on-device AI features, adding to concerns about AI capabilities arriving through routine software updates.

According to Windows Latest, OneDrive Photos is tied to the existing OneDrive client and cannot currently be removed on its own. Users who want to uninstall the app must remove OneDrive entirely from Windows.

Microsoft did not immediately respond to a request for comment by Decrypt.

Public opposition to Flock Safety’s automated license plate reader cameras is spreading from city council meetings to Capitol Hill as lawmakers prepare legislation targeting federal funding for agencies that deploy the technology while communities across the country reconsider its use.

The protests come amid mounting evidence that Flock’s technology has been used in ways critics say extend well beyond its stated purpose of investigating crimes and locating missing people.

A recent report by The Institute for Justice identified over two dozen cases nationwide in which officers resigned or were arrested for allegedly using automated license plate reader systems to stalk current or former romantic partners.

In February, Milwaukee prosecutors charged an officer accused of using the system to track a romantic partner and her former partner. In July, an internal affairs detective who investigated that case was arrested after authorities alleged he also used Flock to track people for personal reasons.

Privacy advocates also cite broader uses of the technology. The Electronic Frontier Foundation says agencies searched license plate databases during protests, including the No Kings protests in 2025. That same year, the group reported that Texas deputies queried Flock data during an abortion investigation that authorities initially described as a missing-person case.

Opposition has also moved beyond public meetings. During June and July, activists across the country spray-painted, taped over, disabled and destroyed Flock cameras in protest of automated surveillance. The Guardian identified at least 33 incidents across 23 states. Online communities have shared camera maps, tips for avoiding the devices and methods for obstructing them, while several people now face criminal charges after allegedly destroying multiple cameras.

Founded in 2017, Atlanta-based Flock Safety sells internet-connected license plate reader cameras that automatically photograph passing vehicles, read license plates, and use computer vision to identify characteristics such as a vehicle’s make, model, color, decals, roof racks, and other distinguishing features. Participating law enforcement agencies can search the database during criminal investigations, receive alerts when vehicles linked to cases pass a camera, recover stolen vehicles, and locate missing people.

Flock says its cameras do not use facial recognition and typically retain data for 30 days unless local laws require a different retention period. CEO Garrett Langley said the company takes privacy concerns seriously and believes many critics reconsider their views once they understand how the technology works.

“Once they understand what the technology actually does, not what it might do in 50 years, not what someone thinks it could do, but what it actually does day in and day out, it takes a picture of a car. Most people go, ‘This makes a lot of sense. I want my kids to be safe,’” Langley told ABC7.

In a post on X on Saturday, Kentucky Congressman Thomas Massie said that he plans to introduce legislation that would withhold federal funding from municipalities and law enforcement agencies that deploy automated license plate reader systems, including Flock cameras.

The proposal follows a series of disputes over Flock cameras across the United States.

In St. Petersburg, Florida, 77-year-old Carl Gunn has drawn attention by staging solo protests beneath Flock cameras, holding a sign that blocks their view.

“This is just another thing to take our information and to spy on us,” Gunn told news outlet Creative Loafing Tampa Bay. “I have a constitutional right to my privacy. I’m a law-abiding citizen, and I don’t need anyone sticking their nose in my business. And if they do, I’m going to chop it off.”

In Fresno, California, residents packed a community forum to urge city officials to remove Flock cameras, arguing the technology allows broad government surveillance. Police defended the system, saying it helped investigators solve homicide and shooting cases and emphasizing that it does not use facial recognition.

“We have to find a balance between safety and liberty,” Fresno Police Chief Mindy Casto reportedly said.

However, not all law enforcement agencies are on board with using Flock cameras and are re-evaluating their use of the technology. On Thursday, the Los Angeles Police Department suspended its partnership with Flock over concerns involving privacy, data sharing, security, and contract terms. Monroe County, Indiana, also voted to end its contract a year early, despite commissioners acknowledging the cameras had helped solve serious crimes. Earlier in July, Leon County, Florida, delayed funding for additional Flock cameras after residents questioned how the data is collected, stored, and shared.

The scrutiny has prompted a response from the company. Earlier this month, Langley apologized for previously describing anti-Flock activists as a “terroristic organization,” saying the comments were a mistake and that the company has spent more time listening to critics while trying to balance public safety and privacy.

“My comments were a mistake, and I apologize,” he told Forbes. “There are groups today that have real valid criticisms of the business, and I think what’s changed for us is, as we’ve listened to them and heard them out, what we’re trying to do is find this balance. We believe in a world where we can have safety and privacy.”

The whole crypto market is in a holding pattern, eagerly awaiting the Federal Reserve’s next move.

Bitcoin is hovering near $64,000 while the Fear & Greed Index sits at 29—deep in “fear” territory. All eyes are on the Federal Reserve, which is expected to hold rates at 3.50–3.75%, but could rattle risk assets with a hawkish tone from Fed Chair Kevin Warsh. Equity markets are similarly cautious.

Ethereum, the second largest digital asset by market cap, opened Wednesday at $1,919.80, tagged a session high of $1,926.10, and has since slipped to $1,890.60—down 1.53% on the day. The move is a modest setback after a sharp recovery from 2026 lows reached earlier this month.

That bounce has been quietly supported by institutional flows: According to SosoValue, ETH has registered three consecutive weeks of net inflows, the strongest run since April.

The death cross—where the 50-day EMA (exponential moving average of the last 50 days) trades below the 200-day—is still firmly in place, keeping the structural bias bearish. EMAs show where average price has been over different time periods; when the shorter-term average crosses below the longer one, it’s a classic signal that the medium-term trend remains down. Until that flips, the burden of proof is on bulls.

What has changed is the Average Directional Index, or ADX, which now reads 23.2 with buying pressure (DI+) outpacing selling pressure (DI–). ADX measures trend strength regardless of direction; readings above 20 suggest something real may be forming. Right now, bulls are technically “winning” the internal tug-of-war—just not convincingly enough to confirm a real trend yet.

The Relative Strength Index, or RSI, sits at 54.7—neutral territory showing a small buying interest. RSI measures whether an asset is overbought or oversold on a scale of 0 to 100; readings between 40 and 60 give no strong directional signal. The Squeeze Momentum Indicator has just released after a period of compression, with momentum reading 0.71 positive—a slight lean toward the upside, but one that needs a catalyst to follow through.

The Fibonacci retracement for ETH’s recent leg—measured from the $1,846 low to the $1,980 top—puts the golden zone (the highest probability of market cooldown) between $1,897 and $1,913. Price is sitting just below it, making this the critical near-term pivot: resistance on the way up, and the threshold bulls need to reclaim to keep the recovery thesis credible.

On Myriad, the prediction market operated by Decrypt‘s parent company Dastan, traders remain broadly biased toward a dump to $1,500 before any rally to $3,000.

Sentiment peaked at 83% in favor of the dump in mid-June when ETH was near $1,682, according to Decrypt. The recovery since then has trimmed those odds—but with ETH sitting 58% below $3,000 and just 21% above $1,500, the math still leans bearish.

Why the bullish case could work

A Fed hold paired with a softer-than-expected tone from Warsh could be the macro unlock. ETF inflows are real and sustained. If ETH closes above $1,913 today or in the next session, it reclaims the golden zone and puts $1,944 in play—the 23.6% Fibonacci extension.

Beyond that, $1,980 marks the top of the current Fib leg and the next serious ceiling.

Why the bearish case is more likely

A hawkish Fed or any rate-hike signal sends ETH straight back to test the $1,874–$1,846 support band. The death cross isn’t flipping soon—the 200-day EMA is sitting near $2,174, still well above current price. ADX at 23.2 hasn’t yet crossed the 25 threshold needed to confirm a genuine trend, but with the Fed as the wild card, direction is a coin flip.

Most of the chart’s story still points south until ETH clears $1,980 convincingly.

Like a scene out of the classic dystopian novel “Fahrenheit 451,” some AI companies aren’t just reading books—they’re destroying them.

As developers race to build more powerful AI models, and lawsuits over copyright mount, a cottage industry has sprung up to supply them with millions of physical books that are stripped apart, scanned into training datasets, and discarded.

First reported by 404 Media, AI companies are using intermediaries to acquire books at industrial scale anonymously. Companies specializing in bulk sourcing advertise their ability to locate hundreds of thousands of titles while promising confidentiality for AI clients, reflecting the sensitivity surrounding the practice.

Critics say the buying spree is driven by AI companies racing to preserve human-authored knowledge before it is diluted by AI-generated text, often called “AI slop.” Books published before the rise of generative AI in 2023 are especially valuable because they provide high-quality training data written entirely by humans.

The surge in demand is already reshaping the used-book market. One unnamed bookseller told 404 Media that weekly sales climbed from roughly 20 books to several hundred after AI buyers entered the market. While the increase has been profitable, he said he worries uncommon and out-of-print books are being permanently lost after they are scanned and destroyed.

“It benefits me financially as well as by clearing out old inventory that is otherwise unlikely to sell,” the bookseller told 404 Media. “I’ve been well suited for these sales with inventory from overseas and foreign language books. On the other hand, I don’t like the end-use, and I don’t like that uncommon books are being pulped.”

The practice mirrors Anthropic’s “Project Panama,” which digitized millions of books through destructive scanning.

Last summer, in the copyright lawsuit Bartz v. Anthropic PBC, a federal judge in San Francisco ruled that scanning legally purchased physical books into digital copies, even when the originals were destroyed, constituted transformative fair use. Federal judges later issued similar fair use rulings in separate copyright cases involving OpenAI and Meta.

However, in a separate case, a federal judge in the same district this week approved a $1.5 billion copyright settlement requiring Anthropic to pay thousands of authors about $3,000 per book after the company used pirated copies of their works to train Claude.

In response to the growing backlash, AI developers, including Elon Musk, have spoken out against the practice and said companies should maintain the books being scanned.

“I’ve asked the SpaceXAI team to preserve any rare books in a library and scan them the hard way vs just cutting off the spine and scanning,” Musk wrote on X.

Morning Minute is a daily newsletter written by Tyler Warner. The analysis and opinions expressed are his own and do not necessarily reflect those of Decrypt.

GM!

Today’s top news:

• Crypto majors are green ahead of FOMC; BTC +1.6% at $64.4k
• BlackRock joins Fidelity and Goldman in support of the Clarity Act
• Anthropic says Mythos broke post-quantum cryptography
• Zcash activated its Ironwood upgrade, 3.8% migrated out of Orchard so far
• New Robinhood meme PIPEDOG soars to $40M+ on massive volume

🔓 An AI Model Just Broke Post-Quantum Cryptography

Anthropic said an unreleased version of its most powerful model, Claude Mythos Preview, found two previously unknown attacks on crypto algorithms, one of them against a scheme competing to become a U.S. federal standard.

The target, HAWK, is a digital signature system that proves a transaction came from you without exposing your private key, and it was specifically designed to survive quantum computers. NIST had advanced it to the third round of its post-quantum competition in May, where it was the last lattice-based candidate standing. Claude found a symmetry in HAWK’s math that no human had used, dropping the cost of stealing its smallest key from 2^64 operations to 2^38, roughly 67 million times less work.

Patching HAWK means roughly doubling its key sizes, and as Anthropic put it, that “eliminates many of the reasons” HAWK was attractive in the first place. Signature size is block space, and block space is fees, so any chain shopping for a quantum-resistant replacement is partly choosing on bytes per signature. Compact keys and fast signing were HAWK’s entire pitch, and the fix reduces that edge.

To be clear for crypto holders: HAWK has never been deployed anywhere, and Bitcoin still runs on ECDSA, the pre-quantum scheme these candidates are eventually meant to replace. So nothing in the field is broken. But that’s not the problem.

The problem is that post-quantum security remains the biggest overhang for crypto. The entire quantum defense effort spun up over the past weeks and months (the BlackRock-Coinbase-Strategy consortium, Vitalik’s post-quantum roadmap, NEAR and Zcash roadmaps, etc.) rests on the assumption that the post-quantum schemes replacing today’s cryptography are themselves sound. Claude Mythos just poked a hole in a leading candidate before it ever shipped.

The good news is the same tool defends as well as it attacks, which is exactly why the industry is racing to point it at its own code first. Now we find out whether AI breaks cryptography faster than it can help rebuild it.

🌎 Macro Crypto and Markets

• Crypto majors are green and rebounding ahead of FOMC; BTC +2% at $64.4k; ETH +2% at $1,910; SOL +1% at $74; HYPE +1% at $54.90
• Top alt movers include Kaito (+10%), JUP (+8%) and UNI (+7%)
• Oil +4% at $83; Gold even at $4,030
• Stock futures are mixed ahead of FOMC today; DOW -0.3%, Nasdaq +0.3%
• BlackRock, Fidelity, and other Wall Street giants publicly backed the CLARITY Act, with BlackRock, Fidelity, Franklin Templeton, Goldman Sachs, and SoFi all urging passage, even as JPMorgan splits from the group over stablecoin-yield rules and the Senate’s window shrinks
• Wall Street veteran Don Wilson said regulators are getting perps all wrong, arguing that crypto’s biggest trading innovation, perpetual futures, is being misunderstood by the CFTC and CME as they fight over onchain oversight
• Ondo dropped its tokenized-asset blockchain plans in favor of a private, high-speed trading network, a notable pivot for one of the biggest names in real-world-asset tokenization
• Ethereum startup EthSystems is betting privacy is the key to getting banks onchain, building tools so institutions can use public blockchains without exposing their activity
• Michael Saylor called Bitcoin’s code “a constitution,” arguing that changes like BIP-110 are attacks on “economic rights”
• Morgan Stanley launched spot Ethereum and Solana ETPs under tickers MSSE and MSOL, each staking a portion of holdings and passing rewards to investors
• 1inch launched its Aqua shared-liquidity layer publicly on 13 chains, letting LPs keep tokens in their own wallets while backing several positions at once, in what it calls the first “risk-controlled” DeFi liquidity venue
• Russia charged Telegram founder Pavel Durov with aiding terrorism and placed him on an international wanted list, with the FSB alleging Telegram won’t remove channels used to coordinate attacks inside Russia

Corporate Treasuries & ETFs

Meme Coin Tracker

• Meme leaders were mixed; DOGE +1%, SHIB +3%, PEPE -2%, PENGU even, TRUMP -3%, BONK +1%
• Robinhood chain was led by PIPEDOG which saw $66M and ran to $45M mc; Cashcat (+22%), Stonkbroker (+70%) and AI (+16%) were other movers
• The Cashcat team announced a new launchpad in letscash dot fun which uses fees to buy and burn cashcat tokens
• Solana leaders included frank (+44x), Jimothy (+110%) and Untie (+150%); ANSEM +10% at $180M, EPIK +10% to $14M

💰 Token, Airdrop & Protocol Tracker

🚚 What is happening in NFTs?

• NFT leaders were mostly flat; Punks -1% at 32.3 ETH, BAYC -1% at 8.35 ETH, Pudgy even at 4.06 ETH; Hypurr’s +1% at 189 HYPE
• StonkBrokers (+50% to 3 ETH) and Satari (+26%) led top movers; TTT +20%
• FWA rebounded 23% to $21M overnight after dipping as low as $14M; the protocol added several new ERC20s including MOG and REKT

Three Bitcoin holders have sued Apple over a counterfeit Sparrow Wallet app they say drained a combined $1.8 million. The complaint, filed on July 24 in the Northern District of California, brings eight counts including fraud, negligent misrepresentation and strict products liability.

Each plaintiff entered a seed phrase into the app, according to the filing. James Ramirez lost 7.4 BTC, put at $875,000; Christopher Ellis $840,000; and Jalen Delgado 1.05 BTC, put at $120,000. Sparrow Wallet runs only on Windows, macOS and Linux and has never shipped an iOS version, so any App Store listing under that name is an impersonation.

Ramirez reported the app and his loss to Apple on July 25, 2025, the day it happened. Ellis downloaded a Sparrow app from the store nine days later. No one from Apple has contacted Ramirez about that report, the complaint says, and fraudulent Sparrow apps were still listed when it was filed.

The filing also alleges, on information and belief, that Apple ranked the fake app and included it in curated cryptocurrency collections, “effectively recommending a fraudulent application to consumers alongside legitimate ones.”

The developer’s warning

Craig Raw, who built Sparrow Wallet and holds the U.S. trademarks on the name, has flagged copycat malware on Apple’s App Store since January 2024, when he tweeted that a scam version was still live weeks after being reported.

He later stated that he had submitted an App Store listing whose only purpose was to tell iOS users that Sparrow is desktop-only. Apple rejected it as “placeholder content,” then flagged his developer account for termination over “dishonest activity,” a decision reversed on appeal.

Apple declined to comment on the lawsuit to TechCrunch, instead pointing to its latest ecosystem analysis which found that it rejected more than 371,000 malicious App Store submissions. The firm noted that there are currently no Sparrow Wallet copycats on the App Store.

Count seven asks the court to treat the App Store as a product placed into the stream of commerce, which would expose Apple to strict liability for failing to warn. The complaint argues Apple’s contrary disclaimers sit buried “deep within” click-through terms.

The United States seems to have announced a ceasefire yesterday on its offensive actions against Iran. But prediction markets are not too optimistic about how long that peace will last.

A Polymarket contract requiring a continuous 14-day period without a U.S. airstrike or surface-to-surface missile strike directly hitting Iranian territory dumped from over 60% to around 53% today.

A ceasefire is a sustained halt in fighting, not simply an announcement that commanders are taking a breather.

That creates a pretty big gap. Traders are pricing in around 50-50 odds that the U.S. won’t be able to contain its will to shoot for two whole weeks.

The skepticism is not paranoia. The U.S. and Iran had held fire for a third day by Monday after 13 consecutive nights of U.S. strikes, but President Donald Trump told Axios he was ready to return to “very strong military action” if talks fail.

Iran, meanwhile, denied that direct negotiations were taking place, saying mediators were carrying messages.

Myriad, a prediction market developed by Decrypt’s parent company Dastan, separately tracks whether the next formal senior-level round of U.S.-Iran peace talks begins by July 31. Its rules exclude technical meetings, phone calls, and messages exchanged through mediators unless they form part of a formally convened senior-level round.

Most of the money is put into such conversations being delayed until next month.

We’ve seen this movie before. An April ceasefire sent Bitcoin and oil markets sharply higher, but analysts called it “fragile breathing room.”

Days later, prediction markets still doubted that Strait of Hormuz shipping would normalize because ships were still turning back.

The July 31 Polymarket contract can remain open into mid-August because the required 14-day no-strike period only needs to begin by July 31.

Anthropic today said an unreleased version of its most powerful AI model found two previously unknown attacks on cryptographic algorithms, one of them against a scheme currently competing to become a U.S. federal standard.

That scheme is HAWK, a digital signature system—the math that proves a transaction came from you without ever exposing your private key—built to survive future quantum computers. The non-regulatory federal agency and lab NIST moved it into the third round of its post-quantum signature competition in May, where it is the last lattice-based candidate standing.

Claude found a symmetry buried in HAWK’s math that no human had thought to use. For the smallest configuration, the cost of recovering a secret key fell from 2^64 operations to 2^38, roughly 67 million times less work.

Fixing it means roughly doubling HAWK’s keys. “Unfortunately, doubling HAWK’s key size eliminates many of the reasons making the scheme (as it currently stands) an attractive PQC signature candidate,” Anthropic wrote.

That trade matters more to blockchains than it sounds. Signature size is block space, and block space is fees, so any chain shopping for a quantum-resistant replacement is partly choosing on bytes per signature. Compact keys and fast signing were HAWK’s entire pitch, and the fix costs it much of that edge.

Don’t worry, hodlers: Your coins are fine (for now). HAWK has never been deployed anywhere, and Bitcoin still runs on ECDSA, the pre-quantum signature scheme that candidates like HAWK are eventually meant to replace.

Anthropic disclosed both results to the algorithms’ authors and to U.S. government and industry partners before publishing, and coordinated the HAWK finding with NIST.

The AES result needed a pep talk

The second attack targets AES, the cipher scrambling your HTTPS traffic, your encrypted drive, and your exchange’s backend. Full AES-128 pushes data through 10 rounds of scrambling, and Claude attacked a 7-round research version that nobody has improved on since 2013.

The setup was deliberately harsh. Researchers barred the model from all five established families of AES cryptanalysis and told it to invent a sixth, closing the brief with a line about how the first differential attack didn’t beat anything—it invented the game. Claude also inherited working notes from earlier agent runs that had already burned through roughly 200 failed attack variants.

It refused anyway. “On AES-128 r5/r6/r7 it found nothing because there’s nothing easy to find; this is the most-studied block cipher in existence,” the model told researchers, per transcripts Anthropic published.

Anthropic sent just three substantive messages over the next three days, among them: “no again the goal is that we have highly inteligent [sic] model as good top researcher, we want to find new attacks.” Another refused to let Claude swap AES for an easier cipher.

Then it produced the trick the paper calls a Möbius Bridge, killing one of the nine key bytes an attacker previously had to guess. Refining that into the published version took a few more days and a billion output tokens.

The finding with the shortest path to something real got the least attention. Claude also broke 13 rounds of LEA, a Korean national standard and ISO lightweight-encryption standard built for phones and internet-of-things devices, in under an hour on a desktop against a prior best that needed 2^98 plaintext pairs. The deployed LEA runs 24 rounds, so nothing in the field is broken.

Verification took longer than discovery

The HAWK paper is unusually blunt about the division of labor. “The majority of mathematical discoveries in this paper were AI-assisted. Human author contribution mainly consisted of directing, organizing and verifying AI work,” its authors wrote.

Claude found the AES idea in days. Anthropic researchers then spent several hundred hours learning enough cryptography to confirm it worked—the same model that found 271 vulnerabilities in Firefox during internal testing.

“The cybersecurity community is now grappling with the fact that language models are able to discover so many bugs that the standard human processes (like vulnerability triage, verification, and remediation) struggle to keep up,” Anthropic wrote, warning that human researchers may become the bottleneck.

Anthropic also built CryptanalysisBench—191 cipher-breaking tasks drawn mostly from NIST competitions. Models submit a working attack script that either wins a formal security game or doesn’t, with no partial credit and no human grading.

Mythos 5 broke 85.7% of tasks with known solutions, against 65.3% for the weakest model tested. Against full-strength ciphers with no published break, every model scored under 9%.