The story involves fraudulent recruitment, compromised devices, internal permissions, transfers across blockchains and a race to trace assets. Its paradox is straightforward: transactions can be cryptographically valid and still be criminal.
The Binance and Mercado Bitcoin incidents occurred days apart. There are reasons to study them together, but not to declare them a single operation. Binance had published a preliminary reconstruction of the intrusion; Mercado Bitcoin's technical root cause remained unclear.
Scope: investigations were ongoing at the editorial cutoff. A connection between the cases, the Brazilian incident's authorship and its root cause had not been publicly established in the sources reviewed. AI-related tools in Binance's account do not establish an autonomous AI attack.
The attack began before the theft
According to Binance's November 3 preliminary report, the intrusion developed through devices, permissions and relationships of trust rather than an attempt to break blockchain cryptography.
In June, a wallet DevOps engineer downloaded a tampered Homebrew installation while looking for an AI coding tool. Homebrew is a package manager used to install software. Security detected the anomaly, replaced the device and rotated credentials. Even so, the company says information about wallet-team personnel had been exposed.
An unsuccessful intrusion can still provide useful reconnaissance. Learning who works on a critical system and who holds deployment privileges can help an attacker plan the next approach.
The fake job interview
In September, another engineer was approached through a legitimate recruitment platform. People posing as HR and company leadership conducted video calls and used convincing profiles. A technical assignment required downloading a repository and setting up an environment on the corporate computer.
The repository contained malicious code. During installation, a legitimate development component was replaced with malware capable of executing commands and communicating with attacker infrastructure. Binance says endpoint protection did not detect it at the time.
The trick was to make a dangerous action look like an ordinary professional task.
Social engineering manipulates context, identities and expectations to induce an action that benefits an attacker. Technical experience alone cannot compensate for a carefully constructed setting. Nor should the analysis stop at blaming an employee: critical infrastructure must assume that devices and accounts can be compromised and restrict the damage.
How a laptop reached the signing service
Around October 20, Binance says the attackers used an internal AI tool to expand access to wallet-team computers and found a credential permitting container deployment.
Containers are standardized environments for running software. The concern here is the authority to deploy code. That privilege can turn a workstation compromise into a production incident.
On October 29, the campaign reached an internal transaction-signing service. This is the mechanism that produces cryptographic authorization for movements accepted by a blockchain. Abusing it can produce a technically valid transfer with an illegitimate business origin.
That distinction matters: a legitimate authorization path can be misused without breaking the underlying cryptography.
When billions began to move
Binance reported unauthorized transfers across networks, involving XRP, SOL, OP, BTC, ETH and other tokens.
Its estimate of $6.34 billion describes the value of affected assets, subject to market prices. It is not a calculation of the net loss after freezes and replenishment. A trail between addresses may be visible on-chain; identifying who controls those addresses requires additional evidence.
A blockchain records a transaction. It does not automatically reveal a beneficiary's civil identity or provide a chargeback button.
EIP-7702 and MPC: why stopping withdrawals was not enough
The investigation described abuse of EIP-7702, an Ethereum feature that lets an account execute delegated code. The technical specification warns that this code has broad access to the account; restrictions depend on the delegated implementation.
The attackers used the compromised signing service to establish malicious authorizations. Pausing the exchange's normal withdrawal system did not automatically invalidate delegations already recorded on-chain.
The report also describes MPC, multiparty computation, which distributes signing among participants instead of concentrating a complete private key in one place. In this incident, the implementation complicated emergency revocation.
The response required a specific revocation procedure. This does not make EIP-7702 or MPC inherently unsafe. It illustrates why protection and recovery need to be designed together, including the scenario in which the authorizing system has been compromised.
The race between tracing and freezing
Many blockchain movements are public. Investigators can follow origins, destinations, timing, amounts and conversions. The FBI attributed the Binance attack to North Korean activity it calls TraderTraitor, publishing associated addresses.
Some centrally issued stablecoins can be frozen under certain conditions. Intervention needs to happen before further conversions or transfers. Native bitcoin has no issuer with an equivalent ability to freeze a wallet.
By November 3, Binance reported approximately $2.85 billion frozen or intercepted, alongside replenishment measures. Traced, frozen, recovered and replenished funds are different categories. A company's use of its own assets to make customers whole does not mean the stolen assets have been returned.
Mercado Bitcoin: another incident four days later
On November 2, Mercado Bitcoin confirmed a security incident affecting hot wallets, used for everyday operations. Its 9 p.m. statement guaranteed customers would not lose assets. Arkham also recorded the exchange's initial claim that no data breach had occurred.
Hot wallets are roughly comparable to an operating cash register; cold wallets provide more isolated reserves. The analogy explains different custody roles, not an absolute guarantee of safety.
Arkham identified a 0.0001 ETH test transfer on Base before larger outflows and reported transfers of ether, stablecoins and bitcoin. Bitquery analyzed ten networks and estimated $48.6 million. Chainalysis reported activity across 17 networks and roughly $49 million. These are different investigation scopes; this article has not reconciled the underlying transactions to produce its own count.
Moving assets across networks does not mean those blockchain protocols were breached. At the cutoff, the exchange had not published a definitive technical root cause. Binance's fake-interview account cannot simply be transferred to the Brazilian case.
The unsettling detail: 36,000 addresses
Bitquery's analysis, updated November 3, identified a subsequent draining pattern involving 36,042 Ethereum deposit addresses, containing about $266,000 in total. Many held small balances. Address counts do not establish a count of individual affected customers.
The amount is small relative to the total, but the operational scale matters. The ability to move funds from thousands of accounts raises questions about signing mechanisms, cryptographic material or central custody services. Those are investigative hypotheses, not the exchange's official root-cause finding.
Bitquery also observed movements after an initial containment announcement. This deserves explanation, but does not establish the state of every internal system at that moment.
Mercado Bitcoin's November 2 update reported roughly $11 million blocked. External trackers published nearby estimates at other times. Methods, coverage and timestamps can explain differences; the figures should not be added together.
What makes the cases comparable
Their timing and transfers across networks justify comparing custody and incident response. They do not establish a common perpetrator. The FBI attributed Binance's case to TraderTraitor; the sources reviewed did not conclusively attribute Mercado Bitcoin's case.
Chainalysis's role also needs to be clear: it reports participating in the investigation and coordination of asset freezes in Brazil. The cited source does not describe a breach of Chainalysis itself.
The blockchain was not the broken layer
A validated transaction can have been authorized through criminal means. Separate three layers: the cryptographic protocol, corporate custody systems, and the people and processes managing permissions.
The protocol checks technical rules. It does not necessarily know the business legitimacy of the decision behind a signature. The reviewed accounts provide no evidence that Bitcoin or Ethereum consensus was broken.
A company can protect keys with sophisticated mathematics while remaining vulnerable in the system that decides when to use them. Network security and custody security are distinct, complementary problems.
Training alone does not solve the human factor
The false recruitment process combined a real platform, plausible people, a familiar task and trust built over time.
Organizations need to anticipate compromised equipment, deceived employees and vulnerable suppliers. Useful architectural responses include isolated environments for external code, limited and temporary privileges, independent validation of critical operations, behavioral monitoring, specific limits for AI agents and real revocation exercises.
The objective is to prevent human trust from becoming unlimited technical authorization.
AI and the engineering of trust
Malicious actors can use advanced tools to research targets, personalize approaches and automate analysis. But a general capability is not evidence of its use in a particular incident. Binance describes AI-related tools in the compromise chain; that does not demonstrate that AI independently planned or executed the attack. The reviewed sources do not establish attacker use of AI at Mercado Bitcoin.
Defenders can also use AI to correlate events and investigate anomalies. The outcome depends on permissions, oversight and containment as well as model capability.
What about banks and personal assets?
Banks have faced cyberattacks for years. It would be misleading to imply that this is new or to assume identical losses across exchanges and banks. Institutions, products and applicable protections differ.
The practical question is: what happens when a trusted identity is compromised? Can an attacker deploy code, approve payments, change data or preserve delegations after a service is paused?
People managing critical systems need to map dependencies, separate duties and test revocation. Individuals need to understand custody arrangements, liquidity and the protections applicable to their assets. Self-custody reduces some intermediary risks while transferring responsibility for keys to the holder.
The next intrusion may begin with a greeting
Science fiction has taught us to fear machines confronting humanity. A closer danger is people using powerful technology to exploit trust.
At Binance, the preliminary account describes fake recruitment, compromised devices and internal privileges culminating in large transfers. At Mercado Bitcoin, custody systems suffered unauthorized movements while investigators traced assets and coordinated freezes. They are distinct cases, and the Brazilian incident's full origin remained unresolved.
Cryptographic engineering cannot compensate for ignoring social engineering and the paths that turn a human decision into digital authorization.
The next crisis may begin with a professional conversation, an apparently legitimate file or a routine meeting. By the time the organization understands what happened, the funds may already have crossed the world.
If this analysis changed how you look at digital security, follow MnzAI Labs for reporting and reflections on AI, software engineering and technology leadership.
Sources and editorial scope
- Binance: preliminary investigation report (Portuguese)
- FBI: TraderTraitor attribution (English)
- Mercado Bitcoin: November 2, 9 p.m. update (Portuguese)
- Chainalysis: investigation and response in Brazil (English)
- Arkham: asset tracing (English)
- Bitquery: analysis across networks (English)
- Ethereum: EIP-7702 specification (English)
Editorial cutoff: November 3, 2026. Architectural recommendations are MnzAI Labs editorial interpretations. Reports and estimates may change.
