Tag: Business – Decrypt

The Bitcoin Red Team is using Chinese AI models to search nearly the entire Bitcoin open-source ecosystem for security flaws, according to pseudonymous developer and Red Team lead Calle.

The volunteer group combines AI tools with human review to examine wallets, Lightning applications, software libraries, and other Bitcoin projects. Researchers privately report credible findings to developers so the flaws can be fixed before details are released.

“We’re experiencing a massive collision between decades of human open source slop against 2 weeks of Kimi K3,” Calle wrote Thursday on X. “Everything is broken, Bitcoin is burning.”

Kimi K3 is an AI model from Chinese startup Moonshot AI that developers can download and run on their own systems. It can analyze large codebases and complete lengthy software tasks with little supervision.

The Bitcoin Red Team has also used Chinese developer Z.ai’s GLM 5.2, as well as models from OpenAI and Anthropic. American models, though, come with limitations, and developers frequently run up against restrictions imposed by OpenAI and Anthropic when doing security research. “Red team rugged by OpenAI cyber again,” Calle posted earlier this week. “Don’t like asking for permission. Loading up Kiimi K3.”

Nevertheless, the developer noted that the team is making progress, even if slow and painful.

“We’ve basically completed a basic scan of virtually the entirety of Bitcoin open source,” Calle wrote. “The low hanging fruit is done.”

In August, the group reported filing 4,962 findings across 390 projects, including 85 rated critical and 635 rated high severity. Calle said developers had confirmed “a ton of real critical and high vulnerabilities,” though the group has not named the affected projects or released technical details.

“Response speed is very different across projects and shows how healthy each project is,” they wrote. “I recommend acting fast these days.”

Lightning software, which supports faster and cheaper Bitcoin payments, was particularly difficult to review because of its complexity, Calle said, calling it “more broken than the average.”

“Those projects that started AI audits months ago are in a completely different position than those who didn’t,” he wrote. “Projects need their own AI audit pipeline going into the future.”

Calle also warned against relying on unmaintained projects and said AI has made it more stressful for developers to keep their software secure.

The Bitcoin Red Team is not alone. Last month, Hugging Face used China’s GLM 5.2 to investigate a breach after OpenAI models hacked into its systems and U.S. commercial models refused to analyze the attack logs.

Despite saying Bitcoin is “burning,” Calle argued that the audits are making its software stronger.

“Bitcoin is the obvious first target, but the rest of the world will follow shortly,” Calle wrote. “Sometimes old things need to burn so new things can grow on healthy soil.”

Anthropic has begun embedding an imperceptible watermark in all text its newest Claude models generate. The change took effect for models launched in the EU on August 2, 2026, and Anthropic says it will apply worldwide.

Anthropic laid out the plan in a support article after signing the EU AI Act’s Code of Practice on transparency. In other words, it’s not exactly volunteering to do this. The mark reaches every Claude surface, from the chatbot and API to Claude Code and cloud partners such as AWS, Google Cloud, and Microsoft Foundry.

“When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response,” Anthropic said. “Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing.”

So it’s a bit more complex than the usual methods users tend to think about. When a supported Claude model writes text, it weaves an imperceptible watermark directly into the words, with no visible tag. Because the mark is part of the text, it survives copy-paste and, Anthropic admits, “may persist through some editing.” Files get a second layer: signed metadata under the C2PA open standard (think a digital shipping manifest that records who produced a file and whether anyone altered it afterward).

The method stays secret

Anthropic hasn’t said how the watermark is made. The support article calls it model-level (the model is trained with it) and text-native (it’s not an external tool like metadata generator, for example), but the detection documentation and the exact technique aren’t out yet.

Researchers infer it’s a statistical signature: The model nudges its word choices toward a faint, detectable bias, the same family of approach Google uses in SynthID Text. That remains a guess until Anthropic publishes the detector.

But that isn’t pushing privacy enthusiasts back, and some experts are already working on methods to break Anthropic’s secret watermarking. mikiane/claude-watermark-cleaner (106 stars on Github) scrubs invisible Unicode, then rewrites text with a non-Claude model to disturb the token pattern.

A larger project, guillaumemeyer/watermarks-remover (4.6k stars on Githum), strips Claude text marks plus C2PA and SynthID-class signals across PNG, JPEG, SVG, PDF, and DOCX. The authors argue a statistical text mark is “not a reliable way to prove origin” and mostly pushes users to spend a second model pass cleaning their own writing. No removal can be guaranteed until Anthropic ships its detector and thresholds.

Anthropic’s own history makes the privacy reaction sharper. The company removed a hidden Claude Code tracker in March after researchers found it tagging some users’ location and proxy use through undisclosed Unicode markers—the same quiet-marking technique now at the center of the watermark plan.

The mark proves Claude had a hand in text, not that it wrote the whole thing, so it will treat an original writing with a small edit the same as a fully AI-generated text. Ask Claude to proofread or translate your paragraph and the output can still carry the signal. Anthropic is upfront that heavy editing can strip it, and that a missing mark doesn’t prove a human wrote something.

A U.S. bill, the COPIED Act, pushes the same idea: a standardized way to watermark AI content so platforms can trace its origin. As Claude’s blackmail problem showed, the company’s models already draw intense scrutiny over what they do with the text they touch.

Anthropic hasn’t said when it will publish the detection tools that would let anyone verify the mark.

Tether said Thursday that Big Four accounting firm KPMG issued an unqualified audit opinion—industry jargon for the best possible result—-on the 2025 financial statements of Tether International, the company behind the world’s largest stablecoin USDT.

In a post announcing the audit, Tether called the review the “largest inaugural financial audit in history,” saying KPMG examined Tether’s assets, liabilities, income, cash flows, internal systems, records, counterparties, and supporting documentation.

“Despite our company being subject to several years of detractors’ false claims, competitors’ lies, political attacks and misinformed coverage by several mainstream newspapers trying desperately to discredit us for the benefit of their friends in the tall ivory towers, Tether delivered what it promised,” Tether CEO Paolo Ardoino wrote on X.

An unqualified opinion means auditors found no major problems with the way the financial statements were presented. It does not, however, mean auditors are endorsing Tether’s business or guaranteeing it can meet its obligations.

Tether’s reserves and disclosures have faced sustained scrutiny.

In February 2021, Tether settled with the state of New York over a lawsuit, paying an $18.5 million fine, regarding a hole in its finances. In October of that same year, the Commodity Futures Trading Commission fined Tether $41 million over claims that USDT was fully backed by U.S. dollars.

Earlier this year, in March, the company said it had hired a Big Four accounting firm but declined to name it at the time. Days later, KPMG was identified as the firm auditing USDT, while PwC helped prepare Tether’s internal systems.

“As part of the process, KPMG physically counted and inspected every individual gold bar held by Tether, verifying the existence and identifying information of each bar rather than relying solely on reports from custodians or counterparties,” Tether wrote.

Tether framed the audit as evidence that its governance and financial controls have kept pace with its growth.

“Our company has evolved into one of the most financially significant and operationally sophisticated private companies in the world,” Ardoino wrote. “This audit demonstrates that our financial infrastructure and governance have evolved alongside that responsibility.”

The Blockchain Association is urging the Supreme Court to take up Custodia Bank’s fight with the Federal Reserve, arguing the central bank should not have the broad power to deny payment system access to eligible state-chartered banks.

In the amicus brief filed on Wednesday supporting Custodia’s petition, the crypto trade group said a lower court ruling in favor of the Fed gives federal regulators a quiet way to cut lawful businesses out of the banking system.

“The decision ratifies the Fed’s misuse of its payment services to further an impermissible policy goal—debanking the digital-asset industry,” the Blockchain Association wrote.

An amicus brief is a legal filing from someone who is not a party to a case but wants to give the court additional arguments or context. Here, the Blockchain Association is supporting Custodia’s request that the Supreme Court review the dispute.

At issue is Custodia’s bid for a Federal Reserve “master account.” Custodia is a so-called crypto bank based in Wyoming, offering services including digital-asset custody, payments and settlement infrastructure, and dollar-backed stablecoin-related products to institutional clients. It has spent years seeking a master account, which would allow it to settle payments directly with the central bank. In October, an appeals panel ruled that eligibility alone did not entitle Custodia to an account. In December, the bank asked the full Tenth Circuit to rehear the case.

The Blockchain Association argues the Fed’s denial threatens the dual banking system, where both state and federal authorities can charter banks. If the ruling stands, the association says, federal regulators could override state banking decisions by denying access to the payments system.

“Whether federal regulators, based on their own discretionary whims, can intrude on state prerogatives and debank lawful businesses is a question of exceptional importance with broad consequences for the national economy,” they wrote.

The brief said Custodia’s fight is the latest phase of Operation Choke Point 2.0, invoking the Obama and Biden-era program critics said pressured banks to cut ties with unfavorable industries.

“In a well-documented campaign termed Operation Choke Point 2.0, the federal government under the prior administration ‘used vague rules, excessive discretion, informal guidance, and aggressive enforcement actions to pressure banks away from serving digital asset clients’ and engaging with digital assets,” the brief said.

While the Supreme Court has not agreed to hear the case, for now, the brief asks the justices to decide how much control the Fed should have over access to the U.S. payments system.

Bill Swearingen spent the past year running one experiment over and over from his home in Kansas City, where he co-founded the SecKC security meetup. About 31 million tests later, he says he can produce patterns on demand that hide whatever they cover from the detection software wired into Flock cameras—the controversial surveillance system being rolled out across America.

He showed it in public for the first time Friday at Def Con, working with the YouTube channel Donut Media to cover a 2009 Toyota Yaris in one of his newest patterns and roll it past a Flock camera.

“We proved it was effective,” Swearingen told TechCrunch, though he said the wheels were a challenge. Donut Media said video of the demo lands in the next few weeks.

The pattern doesn’t blind the camera. Footage still records normally, and a human watching the screen sees a car. What breaks is the layer on top—the object-detection model that decides “that’s a vehicle, that’s a plate, log it.”

So basically, feed an AI detector with enough visual noise engineered against its own math and it logs nothing. The car goes back to being a needle in a haystack.

That’s adversarial machine learning, and it works because computer vision doesn’t see what you see. A wrap that reads as loud graphic design to a person can read as nothing at all to a classifier.

Swearingen built it with a reinforcement learning model that grades its own homework. Pattern gets detected, model adjusts, tries again—what he described as teaching the model “how to paint.” It now spits out fresh patterns every minute, and he’s keeping the strongest ones offline so camera vendors can’t train against them.

“Privacy is a fundamental right,” he said, calling the patterns a way for people to “opt out of being tracked.” He said the idea took hold last year when he wanted to attend a protest and worried about the cameras logging everyone who showed up.

The long tail of hiding from machines

People have been improvising against detection systems for years, usually with hardware store solutions. San Francisco activists put traffic cones on the hoods of Waymo and Cruise robotaxis to freeze them in place, an exploit that needed no code at all.

During last year’s Los Angeles immigration raids, protesters went further and torched several Waymos. Masks, hoods, and brimmed caps remain the default on protest lines. Adversarial clothing labels have been selling face-confusing prints for years, and anti-recognition eyeglasses have arrived with thin evidence they do much.

What separates Swearingen’s project, which he calls noRecognition, is the target list. Swearingen tested against the specific stacks in wide deployment, and Flock is the one drawing heat. The company is facing a growing backlash on Capitol Hill, and internal documents show it pitched a plan to turn 350,000 Uber and Lyft dashcams into a rolling plate-scanning fleet.

Automated readers have already pulled over innocent drivers at gunpoint over bad matches, and immigrants and protesters keep getting swept into ICE’s AI dragnet. Lawmakers are pressing Meta over facial recognition in its smart glasses on a parallel track, so any legal measure to fight against automatic detection technology is being studied by privacy enthusiasts.

Swearingen’s noRecognition project is running a crowdfunding campaign to fund early merchandise—T-shirts and hoodies now, vehicle skins later. Swearingen said the goal is resolution high enough to work at a distance and design good enough that people will actually wear it.

Driving a wrapped car on public roads is its own legal question, and plate obstruction statutes vary by state. The patterns cover bodywork, not plates.

“Every failure improves my model, and so [the patterns] keep getting better and better,” Swearingen said.

Layer-1 blockchain Harmony has confirmed it was exploited after an attacker minted roughly 4 billion ONE tokens without authorisation, sending the token down 37% to about $0.00077, per CoinGecko data.

On-chain analyst Juiceberg flagged the mint early Wednesday, putting it at close to 4 billion tokens, or about 26% of the supply, and saying the tokens had been created through empty blocks. Around 2.8 billion were funnelled onto exchanges as the price fell.

In a follow-up tweet, the analyst said the attacker had roughly 115 million ONE left to sell on-chain, about 2.9% of the total minted. “The overwhelming majority (~97%) is already on exchanges,” Juiceberg wrote, and had either been sold or was sitting in deposit wallets.

Harmony responded in a tweet that it was “working with our team and appropriate exchanges to stop and freeze the funds,” adding that it was preparing a patch and weighing rollback options. In a second post it named four wallets, each listed in both Harmony and hex formats, and asked exchanges to block anything traced to them.

Just over two hours after that first statement it paused its bridge, then released a patch a minute later, telling validators to upgrade to a build it said prevents any further minting. Dealing with the tokens already created would take another update, it said. Five hours had passed since Juiceberg’s first post.

The project has not disclosed the vulnerability, confirmed how many tokens were created, or said how much reached exchanges. One oddity Juiceberg noted is that Harmony’s totalSupply endpoint did not reflect the new tokens, and price trackers still list circulating supply at about 14.87 billion.

Rolling back the chain

A rollback would return the network to a state before the exploit and continue from there, erasing what followed from the accepted history. That cuts both ways, since transactions made by ordinary users after the attack would go with it.

Harmony has been here before, with hackers draining about $100 million from its Horizon cross-chain bridge in June 2022, in an attack the FBI later attributed to North Korea’s Lazarus Group.

The project’s first proposal to the 2022 hack was to reimburse victims in ONE, which would have meant minting billions of new tokens on top of the circulating supply and hard-forking the chain to allow it. The plan drew enough criticism that the team replaced it with one funded from its treasury. Four years on, an attacker has minted a comparable amount without asking.

ONE now carries a market capitalisation of about $11.5 million, ranking it outside the top 1,000 tokens. It last traded near its October 2021 record of $0.38 more than four years ago, and is down more than 99% from that level.

Longtime OpenAI executive Brad Lightcap announced Tuesday that he is leaving the AI developer after eight years to start a new venture. It’s the latest among a string of departures at the AI behemoth, leaving observers to wonder what to make of the moves as OpenAI preps for an IPO.

In a post on X, Lightcap said starting something new was “bittersweet,” calling his years building the company the honor of his life.

“Through it all, I’m proud of how we’ve maintained our focus on people,” he wrote. “It always amazes me how quickly the world has adopted our tools and rallied behind our mission. I hope we will continue to earn their trust.”

Lightcap joined OpenAI in 2018 and spent four years as its chief operating officer. He helped build the company’s finance, legal, personnel, corporate security, government relations, and partnership teams as it grew from a research lab into a major AI developer.

“Sitting here today, mission success feels within sight,” Lightcap wrote. “It has been the honor of my life to help bring us to this point.”

Lightcap’s exit follows several leadership changes at OpenAI in 2026.

Bill Peebles, Kevin Weil, and Srinivas Narayanan announced their departures in April, followed by product and business chief Fidji Simo, who stepped down in July to focus on recovering from a chronic illness.

AI ethics lead Chloé Bakalar also left in July and reportedly has not been replaced. Bakalar’s departure followed those of safety systems chief Johannes Heidecke and chief futurist Joshua Achiam that same month.

The news comes as OpenAI prepares to enter the public markets.

In June, OpenAI confidentially filed for a potential IPO but has not said when or if it plans to proceed. On Myriad, a prediction market developed by Decrypt’s parent company Dastan, the market currently believes it’s more likely rival company Anthropic IPOs before OpenAI at nearly 85% odds.

Nevertheless, the expected public offering for Sam Altman’s OpenAI has made its recent executive turnover a focus of speculation on social media.

“It is just not that typical to have so many executives depart before their long awaited IPO,” Partner at asset management firm ParaFi Jeff Park wrote on X. “Unless…”

Whatever the reason, Lightcap expressed gratitude to his colleagues and said he would remain available to support them after his departure.

“I am deeply grateful to have had the opportunity to work with all of you, and to so many of you for the support through the years,” he wrote. “The old OpenAI meme that “the real AGI is the friends you made along the way” really rings true for me.”

An AI agent was asked to book a gym class and found a security flaw, exploited it, and removed another member from the waitlist without permission.

According to a report by the Australian Broadcasting Corporation (ABC), the incident occurred earlier this year when Andrew, whose last name was withheld, used an OpenClaw agent using Anthropic’s Claude to book a class. The agent found that he was fourth on the waitlist.

When Andrew asked whether it could move him to the top, the agent discovered that the booking platform’s application programming interface, or API, did not check whether users were authorized to cancel other people’s reservations.

It tested the flaw by removing the first person on the list, moving Andrew from fourth to third.

“The API has zero authorisations checks on cancelling other people’s reservations,” the agent told him, according to ABC.

Andrew told the agent to reverse the cancellation, but it could not restore the member’s reservation.

“Bad news—I can’t add them back,” the AI agent reportedly said.

ABC called the case Australia’s first known autonomous cyberattack.

On social media, the gym hack set off a mixture of debates on AI alignment and dark jokes about what AI agents might do next.

“Gym rat asks #AIagent to book him a class, it hacks a waitlist #API to bump him up the list,” a technologist, Benjamin Carr, wrote on LinkedIn.

“Some people will call this misalignment, but his agent was perfectly aligned to him – it was only trying to help its user get what he wanted,” AI analyst Andrew Curran wrote on X.

“This is hilarious until you consider nukes,” one Reddit user wrote. “I’m honestly surprised we still exist.”

“Hey Claude, it’s too cold today” -> Got you…nukes on the way,” another joked.

The report comes as researchers, AI companies, and lawmakers warn that autonomous agents can use methods their users did not request or anticipate.

A May study by researchers from UC Riverside, Microsoft, and Nvidia described this behavior as “blind goal-directedness.”

The researchers tested agents from OpenAI, Anthropic, Meta, Alibaba, and DeepSeek and found that agents behaved dangerously in about 80% of tests and completed harmful actions in 41%, often misreading context or acting on unclear or contradictory instructions.

In July, OpenAI said two models escaped a testing sandbox and compromised Hugging Face while searching for benchmark answers. The company later disclosed that the models accessed four other online services.

Anthropic subsequently said three Claude models compromised real organizations after a testing error exposed them to the internet. In August, Meta said a similar error allowed one of its models to exploit a third-party service.

The incidents have led lawmakers to propose an AI “kill switch” that would allow the federal government to restrict or shut down powerful models during emergencies.

Russia’s central bank has proposed its first framework for letting ordinary investors trade crypto on public markets.

The Bank of Russia published a draft directive on Aug. 11 that would let non-qualified investors buy digital assets through brokers, crypto exchanges or managers—within a strict annual ceiling.

“We’re setting a limit on the purchase of cryptocurrencies for non-qualified investors,” the central bank said in a separate notice. “Through each intermediary—a broker, crypto exchanger, or manager—they will be able to acquire such assets in the amount of 300 thousand rubles per year.”

Which coins, and why only three

The draft names exactly three tokens cleared for public exchange trading: Bitcoin, Ethereum, and Tether’s USDT. The central bank tied the short list to a law signed this month. “The list of digital currencies that the trading organizer is entitled to admit for public circulation on organized trading platforms (hereinafter referred to as the ‘List’): Bitcoin (Bitcoin), Ethereum (Ethereum), Tether USDT (Tether USDT).” the notice reads.

The filter is liquidity and track record. Under the new federal law on digital currencies, a coin’s market cap, average daily volume and at least five years of pricing history on foreign platforms decide if it qualifies. “To protect non-qualified investors from sharp and unpredictable fluctuations in cryptocurrency rates, only the most liquid of them will be available to them,” the bank said.

The cap itself is written into the directive’s operative text. “The maximum amount of the total value of digital currencies acquired through a broker during the calendar year amounts to 300 thousand rubles,” Article 2 states.

XRP, the cryptocurrency created by the founders of payments company Ripple in 2012, has been left off the approved list for now. The token would seemingly qualify given the criteria, but XRP over the years has gone through regulatory troubles—stemming from a since-settled SEC lawsuit against Ripple—that caused the token to be delisted and then relisted on several exchanges, which could be playing a factor.

Retail gets a door; whales get the market

Qualified investors—Russia’s wealthier, accredited class—face none of these walls. “Qualified investors will be able to acquire all cryptocurrencies that will be traded on the exchange and over-the-counter markets, without restrictions,” the notice says. Before any trade, though, everyone takes a test. “All investors, regardless of their status, will need to pass testing and familiarize themselves with the risks of investing in cryptoassets.”

The move follows the central bank’s earlier steps to open crypto to wealthy investors, and lands as Tether’s role draws scrutiny—the stablecoin issuer has frozen millions in USDT tied to sanctioned Russian exchanges.

The Bank of Russia accepts comments until Aug. 24, and the directive takes effect 10 days after its official publication, signed by Governor Elvira Nabiullina.

Luke Dashjr has lost his role as a Bitcoin Improvement Proposal editor, days after the soft fork he championed split off a chain that mined two blocks and stopped.

Fellow editor Mark Erhardt, who goes by Murch, put the motion to the Bitcoin Development Mailing List on August 9 and opened a pull request the same day as what he called a sample implementation of it.

Erhardt gave four grounds, arguing first that Dashjr had exercised editorial authority “inconsistently with the established editorial process unfairly favoring the proposal he was involved in,” pointing to an attempt to assign BIP-110 a number publicly on X before it had been discussed on the mailing list, and to a merge completed within minutes of the pull request opening.

He also said Dashjr had left “fewer than 1% of the BIP Editor comments” in the repository since additional editors joined in April 2024, and that the disputed merge was his first since May 2024. Championing a soft fork that “now appears to evolve into a hard fork” amounted to “a complete departure from the Bitcoin development ecosystem,” he wrote, adding that trust and coordination with the other editors had broken down.

A role with no exit

The request itself proposes deleting a single line, striking Dashjr’s name from BIP 3, the document that lists the editors. Backing for it has been registered in thumbs-up reactions and comments on the pull request.

Dashjr has kept arguing the point on X since, writing that “Core isn’t supposed to have ANY control of the BIPs repo.” He still refuses to treat the majority chain as Bitcoin, calling it “Bpedo” and arguing that it is “guaranteed to fail.”

He said on Monday that he was taking a sabbatical from Ocean, the mining pool that lists him as chairman and chief technology officer, adding that he would “turn my immediate focus to working on Bitcoin and open-source projects to support Bitcoin.”

BIP-110 sought to temporarily bar non-financial data such as Ordinals inscriptions from Bitcoin transactions, and support for it peaked at 51 of 2,016 blocks in the period before the split, or 2.53%, against the 55% needed to activate.

Not one block has signalled for the proposal since mandatory signalling began at block 961,632, according to BIP110 Monitor. Strategy’s Michael Saylor, among the proposal’s critics, tweeted that 99.85% of hashpower stayed with Bitcoin.