How it works
What happens to your jar
- Unpack.
The archive is opened in memory with limits on size, entry count and compression ratio (zip-bomb protection). Entries with path-traversal names are flagged and never extracted. Jars inside jars are scanned too. - Static analysis.
Every class file is parsed at bytecode level. We look for network use, token and credential access, process execution, runtime class loading, backdoor patterns, obfuscation and known-bad signatures, and record the exact class, method and constant. - Network and string analysis.
URLs, domains, IP addresses, webhooks and emails are extracted, compared with a block list and, when possible, with the official Modrinth / Hangar listing. - Behavior sandbox.
The jar is run in a throwaway, offline sandbox with fake accounts, tokens, browser cookies and wallet files at the real paths. All network traffic goes to a local sinkhole that records it. If a decoy is read, or a fake token leaves the box (plain, base64, URL-encoded, hex or hashed), that is evidence of theft. - AI review.
Nimbus AI explains the findings in plain English. Everything from the jar is treated as untrusted data, and the AI cannot decide the verdict. - Following referenced files.
If the jar points at other files (a download URL, a paste or raw-file link, or an address it contacted in the sandbox), our worker fetches them with strict safety limits and scans them as data, including images that hide programs inside them. Fetched files are hash-checked on VirusTotal. Nothing is ever executed. Every source checked, and what it said, is listed in the Investigation tab. - VirusTotal backup.
If our own analysis is unsure, or the AI review and sandbox could not run, we also check VirusTotal using only the file's SHA-256. We support VirusTotal for exactly those moments, so a hiccup on our side does not leave you with no answer. - Verdict.
Static findings, sandbox evidence and the AI’s notes are combined by fixed rules into a verdict and a 0 to 100 risk score.
How verdicts are decided
Malicious
- A fake credential planted by the sandbox left the machine (confirmed).
- The jar read decoy credential files and then tried to contact the network.
- It matches the JarScan block list or a YARA rule marked malicious.
- One class both reads credentials (tokens, browser data, wallets, Discord tokens) and has an outbound channel.
- It contains Discord token patterns together with a network channel, or persistence plus shell commands plus network in one class.
Suspicious
- Any high-severity behavior (for example running OS commands, a Discord webhook, hidden OP commands).
- Three or more medium behaviors, or archive anomalies.
- Obfuscation indicators. Obfuscation alone is never Malicious.
- Network attempts, spawned processes or files written outside the game folder in the sandbox.
- The AI raised a concern the rules did not (it can only move Clean to Suspicious).
Inconclusive
The analysis was incomplete (corrupt or abnormal archive, mostly unparsable classes, no code) and nothing above applied.
Clean
None of the above. Clean is not a guarantee.
Limits
Static analysis can be evaded by determined authors, and a sandbox run can miss code that waits for a specific date, player or condition. The sandbox is not a perfect replica of a real game. Treat JarScan as one signal among several.