Whitepaper

01

Abstract

Traditional antivirus and EDR alert after a file lands or executes. Web3 professionals lose funds in the second they "just click": a fake recruiting assignment, a malicious GitHub repository, a cracked tool, a second-stage downloader. Preflight moves the decision to T-1: the server fetches the sample, unpacks it, runs static and dynamic analysis against Web3-specific indicators and returns an actionable conclusion — do not open / isolated machine only / low risk but stay careful.

This whitepaper is a project draft. It covers the threat model, detection indicators, architecture, MVP scope and the promises we explicitly do not make, and is updated as the implementation progresses.

02

Background

Attacks on exchanges and Web3 teams have shifted from on-chain contract bugs to people and developer machines. Public post-mortems share the same paths: malicious coding assignments delivered through fake interviews, poisoned developer tooling, an engineer's laptop compromised and used to pivot into wallet infrastructure. In parallel, retail users and operations staff keep running into phishing domains, clippers (clipboard address swapping), malicious browser extensions and unlimited-approval scams.

Existing defences have a timing gap. Local antivirus engages fully only after download or execution. Browser blocklists depend on known bad domains and miss fresh redirects and malicious assignments hosted on legitimate platforms. Generic sandbox reports speak in malware family names and never answer the Web3 question: will it steal MetaMask, will it swap my withdrawal address.

03

Positioning

Category

Web3 pre-click / pre-download safety gate

Not

a general antivirus replacement, an on-chain contract audit platform, an exchange-grade SOC

Core action

paste a link / upload a sample → isolated analysis → risk verdict → then decide whether to open locally

Success

a measurable drop in casual clicks; high recall on known families and typical Web3 credential theft

Explicitly not promised

catching 100 % of 0-days; replacing hardware wallets and privilege isolation

Tagline: drop the link in first, then decide whether to click.

04

Target users

05

Threat model

Preflight prioritises initial access and credential theft that begin with a user opening external content — not end-to-end APT forensics after an intrusion.

06

Detection indicators

Indicators are ordered by how funds are lost. Reports speak to users in "will I lose coins" terms and keep IOCs and the evidence chain for analysts.

A. Entry & disguise

• Install chains demanded by the README: npm i, terraform init, curl | bash, run the installer
• Identity disguise: double extensions, forged code signatures, borrowed IDE / Homebrew / updater names
• Second-stage pulls from GitHub, Vercel, Telegram or file lockers
• Weaponised documents: macros, OLE, external links, HTML smuggling

B. Crypto-specific behaviour (the core)

C. Generic malice that travels with crypto malware

• Persistence: LaunchAgents, Run keys, scheduled tasks, browser hijacks
• Evasion: sandbox / VM / debugger kills, quarantine removal, codesign -s -, process injection
• C2: Telegram bots, GitHub API as a channel, odd DNS / DoH, short-lived domains
• Other: mining, keylogging, screenshots, packaged credential exfiltration

D. User-facing scorecard

07

Architecture

1. 01

   Input

   link / file / repository

2. 02

   Access layer

   auth · quota · upload isolation

3. 03

   Safe fetch

   server-side pull · redirect resolution · object storage

4. 04

   Static analysis

   YARA · manifests · strings · dependency graph

5. 05

   Isolated dynamic analysis

   one environment per sample · controlled egress · destroyed afterwards

6. 06

   Intelligence

   phishing lists · domains · IOCs

7. 07

   AI explanation

   evidence → plain Web3 language

8. 08

   Report

   scorecard · evidence chain · IOCs

08

Pipeline

1. 1

   Submit

   Paste a link or upload a file; optional context ("interview task", "tool from TG") adjusts detection weights.

2. 2

   Safe fetch & normalisation

   Server follows redirects with a cap, verifies MIME and magic bytes, unpacks nested archives, deduplicates by hash.

3. 3

   Static pre-check

   Hash intelligence, YARA, script and lockfile audit, extension manifest permission scoring, entropy and obfuscation features.

4. 4

   Isolated execution

   Windows / macOS / Linux sandboxes by type; monitors wallet paths, clipboard, process tree, DNS / HTTPS.

5. 5

   Web3 rule engine

   Maps behaviour onto scorecard dimensions; a lethal hit short-circuits to High risk.

6. 6

   AI summary

   Conclusion and actions generated only from collected evidence; no evidence-free speculation.

7. 7

   Deliver

   Web report + copyable IOCs; enterprise webhook / Slack; samples never return to the user's machine.

09

Role and limits of AI

What AI does

• Translates technical evidence into Web3 consequences: access to the MetaMask extension directory → you may lose your browser wallet
• Fuses signals: static hits + dynamic behaviour + list results → one scorecard
• Recommends the next action: do not click, switch to an isolated machine, open only on a hardware-wallet machine

What AI does not do

• It does not replace the sandbox: truth about behaviour comes from execution and rules, not from a model "feeling" danger
• It does not promise to recognise every 0-day: adversarial samples may evade execution; the copy must stay honest
• It never touches user keys: the check requires no wallet connection and collects no seed words

10

How it differs

11

MVP & roadmap

Status is updated on this page as the implementation progresses.

12

Security, compliance & abuse

1. Samples are hostile by default

• Per-tenant environments, destroyed after analysis, isolated from the control plane
• Egress allowlist (intelligence and required CDNs) so the sandbox cannot become a bot

2. Abuse & legal

• Rate limits, malicious-upload circuit breakers, no use of the site as a malware relay
• Audit logs retained; data residency optional for enterprise

3. Privacy

• The check requires no wallet connection or seed words; connecting a wallet only shows holdings and unlocks quota, and the sign-in signature grants no permission over funds
• Reports can be shared anonymously at hash level; raw samples are never public by default

4. Disclaimer

• Pre-checks reduce risk; they do not guarantee safety
• High-value operations still belong on hardware wallets and dedicated machines

13

Token & quota

The Preflight token is the utility layer of this site on BNB Chain, not a gate: anonymous checks stay free (a limited number per day). The token launches on flap; this page and the token page update as soon as the contract address is public.

Hold for quota

Daily check tiers by wallet holdings; active after connecting and signing in, reset at 00:00 UTC

Burn for priority

Send tokens to the dead address; every full unit adds one permanent check credit. Once the sandbox is live, burns also buy queue priority and deeper analysis

Verification

The server reads the Transfer(wallet → dead) event from the transaction receipt and credits only the event's from address — no contract interaction, no approvals

What it never does

No custody, no approvals, no signing on your behalf; the token never changes a verdict

14

Business model (draft)

15

Risks & non-goals

• Evasion: samples may detect the sandbox; diverse environments and delayed execution are needed, and limits must be disclosed honestly
• False positives hurt trust: legitimate installers also touch sensitive APIs; verdicts must be graded and appealable
• Cost: one environment per sample is expensive; queues, caching (same hash answers instantly) and graded depth
• Non-goals: no formal verification of contracts, no fund recovery, no replacement for social-engineering training

16

Conclusion

A large part of Web3 security's next battlefield is the moment before the click. Preflight's value is turning isolated analysis and crypto-specific indicators into a gate anyone can use before opening a link.

Next: build the MVP-1 fetcher and Windows / Linux sandboxes → measure recall on public stealers / clippers and simulated interview repositories → open a small private beta.