Binance loses $6.3 billion in the largest crypto theft on record

Binance has confirmed that about $6.3 billion in crypto left its wallets on October 29. Most of the funds moved inside a 20-minute window that began at 20:05 UTC. The theft is roughly four times larger than the $1.5 billion Bybit hack of February 2025, previously the biggest on record. Binance has suspended deposits and withdrawals, says customer funds are safe and has set aside more than $6.34 billion of company money to cover losses.

Blockchain intelligence firm TRM Labs attributes the attack to TraderTraitor, the North Korean unit the FBI also blamed for Bybit. TRM’s reconstruction describes an Ethereum hot wallet handed to an attacker-controlled sweeper contract through 50 pre-signed authorizations. They were batched so that at least one would match the wallet’s moving transaction count. Wallets on nine other chains sent funds through ordinary signed transfers in the same period. Those wallets use different key types and signing schemes, so TRM leans toward a compromise of Binance’s hot-wallet signing system rather than one phished employee. Binance has not published a root cause, and its co-CEO said the company has found no evidence of a private-key leak.

Stablecoin issuers and one network moved faster than the attacker could convert everything. Tether, Circle, Ripple and World Liberty Financial froze about $2.67 billion in stablecoins. Arbitrum's security council forced out another $30 million in ether and ARB, bringing the total to about $2.7 billion by 00:30 UTC. The attacker still holds more than 420,000 ETH and 18,270 BTC, worth roughly $2.4 billion. It sold USDT to market makers at a discount above 10 percent to escape assets that issuers could freeze.

Featured source: The Verge , TechCrunch , TRM Labs analysis .

Other Stories

Leaked DHS records expose 41,000 staff and undercover operations

A leak of Department of Homeland Security records has exposed details on more than 41,000 employees and contractors. The files cover DHS itself and seven agencies, including ICE and Customs and Border Protection. They carry work email addresses, phone numbers and office locations. They also name thousands of internal permission groups that reference undercover operations, wiretap programs and surveillance tools. Distributed Denial of Secrets published the archive and says it came from DHS’s Okta login system, which Okta says shows no sign of a breach. DHS says it is investigating.

Filed from: Ars Technica .

Trump orders talks on foreign-made factory robots, with tariffs held in reserve

A presidential proclamation issued on October 28 declares that imports of industrial robots and machinery threaten national security. It directs the United States to negotiate limits with trading partners within 180 days, after which the president could impose tariffs, quotas or fees. American factories buy most heavy robotic arms from FANUC, ABB, Yaskawa and KUKA, all headquartered abroad. The proclamation itself concedes that domestic suppliers cannot meet demand. An earlier FCC restriction on foreign-made robots exempted the fixed industrial arms this order targets.

Filed from: Ars Technica , presidential proclamation .

Wikipedia’s owner asks court to split AI training from AI regurgitation

The Wikimedia Foundation has filed an amicus brief in the copyright suits that publishers including The New York Times and Dow Jones brought against OpenAI and Microsoft. It argues that using copyrighted works to train a model should generally count as fair use when the output adds something new. It says reproducing material in response to a query should require a license when the answer replaces readers’ engagement with the original. The foundation calls Wikipedia one of the most important sources of training data. A blunt ruling either way, it warns, could cut off the AI tools its editors use or let companies ignore licensing terms.

Filed from: The Verge .

Mistral releases the weights of its trillion-parameter frontier model

Mistral has published the weights of Mistral Large 4, a trillion-parameter model it calls Le Chonk. The files are available on Hugging Face, and the model can also be used through Mistral’s API. The company pitches the release at governments and enterprises that want frontier capability for cyber defense without relying on a vendor that can withdraw access. Mistral says the model matches the open-weight Chinese system GLM-5.3 on cyber tasks while using 47 percent fewer tokens, and that it leads Artificial Analysis’ coding-agent index. Those comparisons are the company’s own.

Filed from: TechCrunch .

Reddit will push visitors to log in as search referrals fall

Reddit said on its third-quarter earnings call that it is moving toward a primarily logged-in experience, with tests outside the United States starting in the fourth quarter. The company wants to protect its data from AI scraping and license it to model providers instead. Reddit reported 130.3 million daily active users worldwide, flat on the previous quarter. Its United States total fell to 50.7 million, down 2 percent from a year earlier. Search referrals dropped by more than a million a week on average in October. Reddit warned that daily users could fall by a low-double-digit percentage next quarter. Revenue rose 53 percent to $895 million.

Filed from: TechCrunch , Reddit results .

DoorDash says shopping agents have sent it only a few hundred orders

DoorDash told analysts on its earnings call that third-party shopping agents had placed several hundred orders in total. The agents named included Muse, Instinct and Fo. For comparison, DoorDash handles about 3 million deliveries on an average day. Chief executive Tony Xu said agents remain a small and fast-moving area, and that the company wants working relationships with the developers behind them, including the possibility of advertising inside agents. DoorDash runs its own text, command-line and chatbot ordering tools, plus an MCP server for corporate accounts.

Filed from: TechCrunch , DoorDash results .

From the Community

An open format for writing requirements that agents can follow

A team at Wack has released SLCR, short for Sliceable, Linked, Checkable Requirements. It is an open format for specifications that both people and coding agents read. The group says it built the format after its own agents drifted from the spec in long-lived codebases, and that SLCR lets teams check whether generated code still matches what was asked. The release includes a command-line tool, a getting-started tutorial and an agent skill, with the source on GitHub. The claims come from the format’s authors, and there is no independent evaluation of the tooling yet.

Filed from: SLCR announcement , Lobsters discussion .