CYOPS · Cyber Operations

The Appliance Was the Way In: Bitget’s Loss, and Four Different Numbers

A zero-day in a third-party security product, a database password read from an environment variable, and a web shell on a second appliance. The transaction-signing control was never broken; it signed requests that already looked authorised.

What happened, in order

Bitget’s first notice is precise about detection and silent about everything else: “At 18:31 UTC on September 24, 2026, Bitget’s security systems detected unauthorized transfers from some of our hot wallets”, with exposure limited to “a portion of the hot wallet and warm wallet layers” and cold wallets “fully secure” [1]. It added that the company “will not speculate on the attack vector until the investigation is complete” [1].

The investigation, carried out with outside firms, later produced a sequence. The earliest malicious activity in the logs dates to 31 August, when a service on one node of a third-party security appliance was hit by a zero-day; further hidden-script activity appeared on two other nodes between 23 and 25 September [4]. At 16:07 UTC on 24 September the attacker reached a second appliance’s management platform using a stolen employee identity; malicious programs started at 17:49 UTC; the first verified theft began at 18:31 UTC; reconciliation flagged a discrepancy at 19:05 UTC and risk controls blocked withdrawals; transfers ended at 21:23 UTC [4].

The most useful number here is not a dollar amount. Initial access was 31 August; theft was 24 September. Roughly three and a half weeks passed inside the estate, on security appliances, undetected — and what raised the alarm was financial reconciliation, not security monitoring.

The entry point was a security product

Bitget says the attackers “exploited a zero-day flaw in third-party security products” [3], and in its own summary describes “a vulnerability in a third-party security product to obtain high-level internal credentials”, with “private-key compromise” ruled out [10]. The products are referred to only as “Product A” and “Product B”; no vendor is named and no CVE has been published [4].

MITRE classifies this as a trusted relationship, where adversaries abuse “an existing connection that may not be protected or receives less scrutiny than standard mechanisms of gaining access to a network” [7]. A security appliance is the extreme case: placed where it can see and reach a great deal, trusted by construction, and often exempted from the monitoring applied to general-purpose hosts because it is itself a security control.

The question is not which vendor this was, but which appliances in your estate can reach your most sensitive systems, and whether anything would notice a shell under their service processes.

A password in an environment variable

The step from code execution to data is the part most worth copying into an internal standard. The attacker “ran a hidden script under the service process, launched a command to read the environment variable containing the database password, and connected to the database” [4].

CWE-526 covers a product that “uses an environment variable to store unencrypted sensitive information”, and gives the consequence: such information “can be accessible by other processes with the execution context, including child processes that dependencies are executed in”, and its contents “can also be inserted into messages, headers, log files, or other outputs” [6].

Environment variables became the default because they beat secrets committed to source control. They are not a boundary: any code that attains the service’s execution context reads them without a privileged operation, without touching disk, and without an event a file-access rule would catch. Hence the investigators’ recommendation to “Remove secrets from environment variables where you can”, using a vault with short-lived credentials [4].

The second appliance, and a borrowed identity

The move onward used two ordinary techniques. The second appliance’s management platform was accessed with an internal employee’s identity [4][5] — what MITRE records as abusing “credentials of existing accounts”, with detection resting on “anomalous logon patterns, abnormal logon types, and inconsistent geographic or time-based activity” [9].

Then “the threat actor deployed a web shell onto the security appliance B and established a Command-and-Control (C2) connection”, before moving “laterally to Bitget’s production wallet job server” [4]. The standard detections for that technique look for unexpected file creation in web directories followed by web server processes spawning shells or interpreters [8].

Neither technique is novel, and that is the point. The novelty was confined to the zero-day at the front; the rest was the familiar sequence, on hosts where the familiar detections were absent.

Four numbers, and why they differ

The loss has been published four ways. Bitget’s first notice gave “approximately $351.6 million” [1]. Coverage of the same period recorded $390.06 million transferred to attacker-controlled addresses [2]. A later account gave “about $387.5 million”, noting the change came from “reclassifying transactions, not from a second theft” [4]. Bitget’s retrospective states the “final verified amount affected is approximately $388 million” across “12 wallet addresses”, and that this “does not represent additional unauthorized transfers following containment” [10].

The spread is not evidence of concealment. Gross outflow, net loss after reclassification and a final verified figure are three different measurements, and an estimate made within hours is a fourth. Record the basis alongside the number: a brief saying “$351.6 million” and one saying “$390.06 million” are both accurate about different things.

What the signing system proves

The most instructive line from the investigation: the signing system “did what it was built to do: sign requests that looked authorized. The compromise happened earlier in the chain” [4]. Bitget’s earlier framing matched — the attacker “compromised a critical backend system within our wallet infrastructure, used it to spoof transaction data, and triggered our authorization process to move funds out” [2].

This is the generalisable finding, and not specific to cryptocurrency. An approval control that validates the form of a request and takes the requesting system’s identity as given is only as strong as the weakest system permitted to request. Hence: “Add approval steps that don’t depend on a single system being trustworthy” [4]. Elsewhere the request is a payment file, a configuration push or a deployment, and the control has the same blind spot.

Attribution: what rests on what

The attribution is reported consistently and hedged differently by speaker. Bitget assessed that “based on IP behavior patterns and on-chain analysis, the attack method in this incident is highly consistent with known patterns of North Korean hacker organizations” [2], and researchers linked the theft to a known cluster through laundering infrastructure overlapping earlier attributed thefts [2]. Elsewhere the company is recorded declining to “speculate on attribution without verified findings” [4].

Both can be held at once, and a brief should. “Highly consistent with known patterns” is a similarity judgement about method and laundering, not an identification. The firms engaged have been named: a third-party investigation involving Mandiant and SlowMist, with wallet tracing cited from Elliptic and TRM Labs [2][5]. Readers needing attribution for a decision should note which part is on-chain clustering and which is behavioural inference.

What to do

  • Map what each security appliance can reach, and treat that as its blast radius, not its coverage [4][7].
  • Remove secrets from environment variables where the platform allows, and move to a vault issuing short-lived credentials [4][6].
  • Hunt for scripts and shells under appliance service processes, and database connections originating from appliances [4][8].
  • Alert on management-console logins at unusual times or origins, and require phishing-resistant authentication on those consoles [4][9].
  • Ship appliance logs off the appliance, so the record survives a compromise of the device [4].
  • Add an approval step for high-value actions that depends on no single system being trustworthy, and wire reconciliation discrepancies into the security response path [4].

The durable lesson is about where trust sits, not about an exchange. The organisation had working reconciliation, automated withdrawal blocks and a signing process that behaved correctly, and still lost the funds, because the thing that vouched for the request had been taken three weeks earlier on a device nobody was watching. Appliances bought to watch the estate are part of the estate, and the monitoring applied to them should say so.

Sources

Every R3KONX article cites its primary material. 10 sources, in order of first citation. Links open the original publication.

  1. [SECURITY NOTICE] Bitget exchange hot wallets Incident, September 24, 2026 Bitget · 2026-09-24
  2. Bitget Says Suspected North Korean Hackers Stole $351.6M After Backend Compromise The Hacker News · 2026-09-25
  3. Bitget Confirms Third-Party Zero-Day Behind $387.5 Million Cryptocurrency Theft The Hacker News · 2026-10-05
  4. Bitget $387.5M Crypto Theft: Zero-Day Security Appliance Breach and Wallet Server Lateral Movement WindowsForum · 2026-10-05
  5. Bitget Confirms Zero-Day Behind $387.5 Million Crypto Theft BlockTelegraph · 2026-10-05
  6. CWE-526: Cleartext Storage of Sensitive Information in an Environment Variable MITRE · 2026-10-06
  7. T1199: Trusted Relationship MITRE ATT&CK · 2026-10-06
  8. T1505.003: Server Software Component: Web Shell MITRE ATT&CK · 2026-10-06
  9. T1078: Valid Accounts MITRE ATT&CK · 2026-10-06
  10. Bitget Security Incident: What Happened, Timeline, Impact and Response Bitget · 2026-10-02

Researched and written by the R3KONX analysis desk from the cited primary material. The sequence and technical detail are as published by Bitget and by the firms it engaged, reported through accounts of those findings; the two third-party products are unnamed in every source and no CVE has been published, so the vulnerable software cannot be identified here. Four different loss figures have been published and all four are given with their provenance. Attribution is reported as the assessing parties worded it, including Bitget's own refusal to speculate. Corrections to event@r3konx.asia.

R3KONX 2027

Work in this domain? So does the programme

4–6 May 2027, World Trade Centre Kuala Lumpur. Curated talks, hands-on training and The BattleGrid.