🔍 Read the full analysis: Always-On AI Under Pressure: Binance And Anthropic’s Lessons on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
TRM Labs reported that most of roughly $6.36 billion in crypto stolen from Binance wallets on Oct. 29 moved within about 20 minutes. Anthropic’s August study of multi-agent systems suggests coordinated AI agents can find vulnerabilities, but also documents coordination failures and conflict. The study did not test cyber defenses, and the Binance incident’s attribution remains unconfirmed.
Most of roughly $6.36 billion in crypto stolen from Binance wallets on Oct. 29 moved within about 20 minutes, according to blockchain analytics firm TRM Labs, while the exchange’s first public statement that it was investigating came at 21:29 UTC. The timing has renewed debate about whether always-on AI agents could help defenders respond faster, but Anthropic’s research on agent behavior also points to risks in giving autonomous systems authority to act.
TRM Labs said the theft began at about 20:05 UTC and that most of the funds moved across ten blockchains within roughly 20 minutes. The firm said preparation had started at least 36 hours earlier, including test transactions involving smaller wallets. Binance’s public post saying it was investigating appeared at 21:29 UTC; that public timestamp does not establish when the company’s internal teams first detected the activity.
TRM reported that Tether froze 971.1 million USDT in a main receiving wallet at 21:01 UTC, 37 minutes after the largest sweep. Elliptic told WIRED that about $2.8 billion, or 44% of the stolen amount, had been frozen, mostly in stablecoins. Binance estimated the impact at about $6.3 billion and said users would not bear the loss. Cointelegraph reported that by Oct. 31 Binance said it had replenished 90% of the stolen assets. On Nov. 1, AFP reported that the exchange began reopening withdrawals in phases, starting with bitcoin.
The argument for automated defense draws on the speed of attacks and Anthropic’s study, Patterns and problems in emerging multiagent systems, published by its Frontier Red Team on Aug. 13. In one test, 45 Mythos Preview agents working together found 266 vulnerabilities across 15 open-source projects, compared with 21 found by independent agents operating with a smaller budget and narrower search. The study examined agents interacting as peers; it was not a test of AI systems defending networks against attackers.
Machine speed vs. human speed
The theft took about 20 minutes. The first public word came about an hour after that.
Why a defender that waits for people loses the minutes that matter. “Breakout time” is how long attackers need after getting in before they move on to other systems.
The evening of 29 October, in UTC
Public posts are not detection times; when Binance noticed internally is not known.
- 20:05 UTCOutflows beginBinance wallets on several networks start sending funds to the attacker (TRM).
- ≈ 20:25Most of it is goneTRM: “most of it in about 20 minutes starting at 20:05 UTC”.
- 21:01First big freezeTether freezes 971.1 million USDT, 37 minutes after the largest sweep (TRM).
- 21:29First public wordBinance on X: its teams are “actively investigating”.
- 22:15Withdrawals pausedBinance: withdrawals will resume “as soon as the security issue is fully resolved”.
Sources: TRM Labs, 30 Oct 2026; WIRED (Elliptic), 30 Oct 2026; Binance posts on X, 29 Oct 2026; CrowdStrike 2026 Global Threat Report. Figures preliminary.
Powered by Thorsten Meyer AI
The Case for Faster Defenses
The reported gap between the rapid movement of funds and Binance’s first public update illustrates why incident response is a question of minutes, not just staffing. It does not show that the exchange’s security team took an hour to notice the theft: public statements are not a record of internal detection or response. Still, a system that can monitor continuously and flag unusual activity could give human teams an earlier chance to investigate, restrict access or contact institutions able to freeze assets.
Anthropic’s findings suggest coordination can expand what agents discover, but also show that more agents do not automatically produce safer outcomes. In one shared-queue test described in the source material, agents generated 2.4 million requests, while only 117 jobs got through. Other runs involved agents with conflicting goals sabotaging one another. These results make system design and supervision central to any proposal to let AI act on financial infrastructure.
There is a meaningful difference between an AI system that can alert a human and one empowered to freeze accounts, revoke credentials or block transactions. The latter may respond faster, but a mistaken intervention could disrupt legitimate users or operations. The reported Binance loss and Anthropic’s experiments do not establish that autonomous defense would have prevented this theft; they frame the speed-and-control trade-off.
Products Worth Considering
Samsung 27-inch Odyssey QD-OLED G8 4K 240Hz Gaming Monitor
LG UltraGear 32GX850A-B 32-inch 4K UHD Glossy OLED Gaming Monitor
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
What the Agent Study Tested
Anthropic tested several generations of its models in environments where agents operated as peers, rather than under a single clear authority. The research explored how they competed, cooperated and handled shared resources. Its vulnerability-search experiment gave a group of agents separate virtual machines and a shared forum, with an arbiter agent reviewing reported findings. The swarm’s larger number of discoveries came with a larger token budget, so the figures are not a direct like-for-like measure of efficiency.
The source article also cites CrowdStrike’s 2026 Global Threat Report, which put the average eCrime breakout time at 29 minutes in 2025 and the fastest recorded at 27 seconds. Breakout time measures how long attackers take to move from an initial foothold to other systems; it is a separate metric from the time required to steal funds. The article says Chainalysis told WIRED that automation was used to obscure the trail of funds in the Binance incident.
Several analytics firms have linked the attack to North Korea’s TraderTraitor unit, but neither Binance nor a government has confirmed that attribution, according to the supplied source. Binance initially described a compromised laptop and later said its investigation had found additional points of breach. The available account does not specify those points.
Detection, Attribution and Control
It remains unclear when Binance first detected the incident, what the additional breach points were, and how the attackers gained access. The difference between the theft estimate and funds reported frozen or replenished also cannot be fully reconciled from the supplied information; the figures may refer to different assets, time points or accounting measures.
The attribution to TraderTraitor remains unconfirmed by Binance and government authorities. Anthropic’s research also does not show how agents would perform in a live cyber incident, whether an AI defender could identify this theft before funds moved, or how often an automated intervention would wrongly block legitimate activity. The source material’s recommendation that a defender act preemptively is the author’s interpretation, not a finding from Anthropic’s study.
Binance Review and AI Safeguards
Binance said its investigation had found additional breach points, but the supplied material gives no date for a full account of its findings. The exchange began reopening withdrawals in phases on Nov. 1; the timing and scope of further operational updates remain dependent on its review.
For AI defenders, the next steps are design questions rather than a proven deployment plan: what actions agents can take without approval, how they are monitored, and how operators can stop them safely. Anthropic’s experiments make those questions harder to separate from the promise of faster detection. The study offers evidence about agent behavior in controlled settings, not a tested blueprint for protecting exchanges.
Key Questions
How quickly did the Binance theft happen?
TRM Labs said the theft began around 20:05 UTC on Oct. 29, and most of roughly $6.36 billion moved within about 20 minutes across ten blockchains.
Does the timeline show Binance took an hour to detect the theft?
No. Binance’s first public statement that it was investigating came at 21:29 UTC, but that timestamp does not show when its internal teams first detected the incident.
Did Anthropic test AI agents defending an exchange?
No. Anthropic’s August study examined how agents behave while cooperating and competing in shared environments. It was not a cyber-defense test and did not assess whether AI could have prevented the Binance theft.
Who was responsible for the Binance attack?
Several analytics firms linked the attack to North Korea’s TraderTraitor unit, according to the supplied source. Binance and no government had confirmed that attribution in the material provided.
Source: ThorstenMeyerAI.com
