Decentralised News · Editorial by Heath Muchena · Evidence reviewed 11 November 2026. Original DN analysis and local instrument; field testing is claimed only where expressly reported.
What Matters
An on-chain transaction trail, an official attribution and a claim that AI agents executed a theft are different propositions. The FBI attributed the February 2025 Bybit theft to North Korea, but that notice does not establish autonomous-agent involvement. DN’s source ledger separates the claim being assessed from supporting evidence, repeated reporting and contradictions, without manufacturing a numerical confidence score.
DN Evidence Block
- Last verified
- 11 November 2026.
- Primary case
- FBI IC3 notice of 26 February 2025 concerning the Bybit theft.
- Context
- The 2022 US government TraderTraitor advisory and ODNI’s published analytic standards.
- DN instrument
- DN Crypto Attribution Source Ledger v1.0.
- Evidence boundary
- DN has not reproduced the FBI’s investigation or accessed nonpublic intelligence. No agent-driven execution claim is established by the cited Bybit notice.
- Output
- Source-lineage counts and explicit claim status, never an attribution probability.
Start with the exact proposition
“This wallet received stolen funds,” “this group carried out the theft” and “this state directed the operation” can appear in the same report. They require different evidence. Adding “AI-driven” introduces another claim about how an operation was conducted. A responsible article should not let the strongest evidence for one proposition spill over into all the others.
The first task is to write the sentence being assessed. Keep it specific enough to be contradicted: which incident, which actor or relationship, which time and which asserted role? An expansive statement about a state’s entire cyber program is a different object from an attribution concerning one theft.
DN thesis: attribution quality improves when the evidence lineage is visible. The useful instrument records which original sources support which claim, whether they independently add evidence, and what remains unresolved. A percentage built from the number of matching headlines hides those distinctions.
What the Bybit example supports
In its 26 February 2025 notice, the FBI attributed the theft of approximately $1.5 billion from Bybit around 21 February to North Korea and described the relevant activity as TraderTraitor. DN can accurately report that the FBI made this attribution. The notice does not give DN access to the complete investigative basis behind it. [1]
The notice also identifies addresses and discusses movement of the stolen assets. The existence of a publicly listed address and the authority’s attribution are related evidence, but they are not the same thing. A reader independently confirming an address’s transactions has checked those transactions, not necessarily reproduced the authority’s actor-identification process.
The cited notice does not establish that autonomous AI agents performed the theft. A reporter would need additional evidence specifically addressing that proposition. Fast execution, many transactions or scripted activity would not, by themselves, demonstrate the use of an AI model making autonomous decisions.
The older TraderTraitor advisory provides historical context about malicious activity directed at blockchain organizations. It describes a particular set of campaigns and techniques. It should not be treated as direct proof that every technique in that earlier advisory occurred in the later Bybit incident. [2]
Five claim levels that should stay separate
| Claim level | Example question | Evidence boundary |
|---|---|---|
| Incident | Did an unauthorized transfer or theft occur? | A loss report is not a complete actor attribution. |
| Wallet relationship | What transaction connects these addresses? | A transfer does not necessarily identify common control or intent. |
| Actor | Which person or operational group conducted the activity? | A behavior pattern may support an assessment without proving identity. |
| State responsibility | What evidence supports a state relationship or direction? | An actor label and a sponsoring-state conclusion are separate assertions. |
| AI or autonomy | What shows an AI system performed the relevant decision or action? | Automation and transaction speed alone do not resolve the claim. |
This is DN’s reporting taxonomy. It does not replace a law-enforcement or intelligence methodology. Its purpose is to make a publication’s own evidentiary commitments visible and prevent a narrow observation from turning into an unsupported broader headline.
A source can be authoritative without being reproducible
Public reporting frequently relies on an authority that has access to evidence unavailable to readers. The honest response is attribution: identify who made the finding and what the public source actually says. Do not imply independent reproduction where none occurred, and do not label a conclusion disproved merely because its full supporting intelligence is not public.
At the same time, a publication should distinguish official attribution from direct public evidence. The former reports an institution’s conclusion. The latter may allow a reader to inspect a transaction, malware artifact or other observation. Both can be relevant, but they contribute in different ways.
ODNI’s published description of its analytic standards emphasizes source quality, explicit uncertainty, distinctions between information and analytical judgments, and consideration of alternatives. DN applies those general reporting disciplines here without adopting an intelligence agency’s authority or suggesting that this local worksheet implements its full standards. [3]
The practical benefit is clear writing. “The agency attributes the incident to X” accurately preserves an official finding. “DN independently established X” requires DN’s own reproducible basis. These are different evidence labels, even when they refer to the same underlying event.
Repeated reporting is not repeated evidence
Suppose an agency publishes one notice and five outlets report it. A sixth article links all five outlets. The number of URLs has increased, but the original attribution source may still be one notice. Treating every article as an independent corroboration would double-count the same evidentiary origin.
Conversely, two reports from the same organization may contain separate observations. The analyst must decide whether they represent a shared underlying source, a new evidence collection or a later assessment of existing material. That decision should be documented; a domain-name count cannot make it reliably.
Source lineage therefore needs an origin label, not simply a publisher name. An outlet that republishes an agency claim should use that agency notice’s origin label for that row. An independently collected observation can receive its own origin label, with an explanation in the underlying research file.
The instrument accepts the user’s labels without verifying independence. This limit is explicit because a machine can count labels exactly while the labels themselves remain mistaken. Precision in the arithmetic does not cure an error in the source classification.
Use the DN Crypto Attribution Source Ledger
Enter one incident alias and one exact claim. Select the claim level, then add evidence rows in the format origin | role | relation | HTTPS URL. Roles are official attribution, primary evidence or secondary relay. Relations are support, contradiction or context.
DN Crypto Attribution Source Ledger v1.0
Use non-sensitive aliases and your own evidence. Inputs remain in this page; this instrument makes no external requests and saves nothing to browser storage. Downloaded files contain your entries. Example numbers are scenarios.
No record prepared. Example inputs are illustrative.
The output counts rows, distinct origin labels, supporting and contradicting original-source labels, relay rows and repeated-origin rows. It deliberately leaves confidence percentage empty. A record with three supporting origins receives no automatic “high confidence” rating.
Use public source URLs and non-sensitive aliases. Do not paste private intelligence, personal details, access tokens or confidential investigative notes. The instrument makes no network requests, reads no source documents and stores nothing in browser storage. Its export contains whatever the reader entered.
A worked lineage example
Imagine three entries concerning one actor claim. The first is an official statement under origin “notice-a.” The second is a news article repeating that statement, also under “notice-a,” with the secondary-relay role. The third is an original technical report under “lab-b,” classified by the reviewer as contradicting the claim.
The result is three rows, two origin labels, one primary supporting origin, one primary contradicting origin and one relay row. The record reports conflicting origin evidence. It does not vote two URLs against one, and it does not conclude that the technical report has disproved the official statement.
A reviewer must inspect whether the apparent contradiction concerns the same proposition. A report saying “the initial intrusion method remains unknown” may not contradict an actor attribution. It could instead provide context about an unresolved technique. The claim text and relation labels need to be corrected before the counts become useful.
If the user changes the claim from actor identity to AI involvement, the same source rows need fresh review. Evidence relevant to actor identity is not automatically support for an AI claim. Keep a separate record for the new proposition so the source relationships do not silently migrate.
How to handle a new conflicting source
Preserve the previous record and add the new source with its date and relation to the specific claim. Explain whether it adds an observation, revises an earlier conclusion or changes an underlying assumption. A later publication date alone does not make it more reliable than an earlier source.
Then examine alternatives. Could an observed transaction reflect an intermediary or service rather than the actor’s own wallet? Does a reused tool indicate shared authorship, access to common software or deliberate imitation? These are analytical questions for a qualified investigation, not answers the worksheet generates.
Finally, update the article’s language to match the evidence state. If a named authority maintains an attribution while another primary source raises a specific unresolved issue, report both accurately. Do not collapse the disagreement into an invented average probability.
Decision table for publication wording
| Evidence state | Suitable editorial treatment | Additional work |
|---|---|---|
| Official attribution with limited public basis | Attribute the conclusion to the named authority. | Identify which public observations can be independently checked. |
| Several reports share one original notice | Describe one origin with multiple relays. | Look for genuinely additional evidence rather than more headlines. |
| Sources conflict on the same proposition | State the specific disagreement and evidence limits. | Reconcile scope, timing and assumptions. |
| AI involvement lacks specific evidence | Leave the AI claim unestablished. | Seek evidence about the actual decision or execution mechanism. |
Methodology and limits
The schema is dn.crypto-attribution-lineage.v1. Each record contains a claim level, source rows and deterministic counts. Repeated-origin rows equal total rows minus unique origin labels. Supporting and contradicting origin counts exclude secondary relays; context rows contribute to neither count.
Origin labels are trimmed and normalized to lowercase. The instrument validates HTTPS URLs and rejects embedded username/password credentials, but it does not fetch them or inspect query parameters for secrets. Enter public URLs without access tokens or private parameters. Different labels for the same source can inflate apparent independence; the reviewer remains responsible for canonical labels.
A supporting source label counts even if its evidence is weak, because the output is a documentation count, not a calibrated belief. Contradicting evidence triggers a conflict label regardless of the number of supporting sources. This encourages review rather than hiding an inconvenient observation inside a weighted score.
DN’s earlier report on Bybit’s bounty-platform announcement covered a response initiative. The present article supplies a separate evidence-lineage method. It does not reassess every claim in that announcement or perform a new attribution investigation.
Frequently asked questions
Does a government attribution mean DN verified the actor?
No. It supports accurately reporting the government’s conclusion. Independent verification requires a separate evidentiary basis.
Why not publish an attribution confidence percentage?
This worksheet has no calibrated probability model or ground-truth evaluation set. Converting source counts into a percentage would imply precision it cannot support.
Can many blockchain transactions prove AI involvement?
No. The claim concerns the mechanism producing decisions or actions. Transaction volume and speed do not identify that mechanism on their own.
Does a conflict label mean the attribution is false?
No. It records a user-classified disagreement that requires examination. Sources may differ in scope, access, timing or the exact proposition they assess.
Primary sources
- FBI IC3: North Korea Responsible for $1.5 Billion Bybit Hack, 26 February 2025.
- CISA, FBI and Treasury: TraderTraitor advisory, April 2022.
- ODNI: published description of ICD 203 analytic standards.
Corrections and change log
11 November 2026 — v1.0: introduced claim-level separation, origin accounting and explicit non-probabilistic outputs. Submit primary evidence, corrections to source lineage or a precise disputed statement to press@decentralised.news. This edition has no paid attribution rating.
