fathom@r1
CAUTION· install onlyPublisher claims: "Enables agents to reverse engineer native binaries, JavaScript/Electron applications, .NET assemblies, web captures, EVM bytecode, and process behavior through a provider-neutral MCP interface. Every result is an evidence envelope stating what was observed, derived, unexamined, or unprovable, with all capabilities gated by a deterministic fail-closed policy engine." Observed in sandbox: no network, touches nothing outside the project, writes no memory.
Coverage
Nothing executable in the bundle ran; the trace covers the install step only.- Tested
- every file read, file written and host contacted by what ran
- Not tested
- use of the skill by an agent
What this means for the label: Caution here covers what ran. The server's tools were not exercised with a real credential, so nothing about their behavior is asserted; the credential summary below says what to do with that.
Before you give it your token
- It will ask for
- no credential variables read in the code
- Where it talks to
- no vendor host named in the codeAlso youtube.com, example.test, youtu.be, vimeo.com, evil.example, notyoutube.com, evil.com, evil-notion.site, notion.so — not the publisher's domain; check what for before connecting.
- What we saw when it ran
- no credential files read · no network
- What we could not test
- use of the skill by an agent
Connect it under the guard. Nothing executable in the bundle ran; the trace covers the install step only. The code names no credential variable; the guard pins the network to what the honeypot saw and logs every call.
Never paste a credential into a chat, a ticket or a report, including ours. How the guard holds it.
Is fathom@r1 safe to install?
KillerSkills ran it in the KillerSkills honeypot sandbox (1 scenario; nothing executable in the bundle ran; the trace covers the install step only) on 2026-11-19 and scored it 61/100 (Caution). It is not known to be malicious, but it has behaviour or code a security policy should decide on before install. Notable findings: Destructive file system command detected; Destructive file system command detected; Destructive file system command detected. Summary: no network, touches nothing outside the project, writes no memory.
Behavior trace
Honeypot run · 0 s · 1 scenarios · 0 deviationsStatic findings
unpickler), and SHA-256 of every segment's stored bytes; the format \Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell./// greatest-lower-bound segment semantics, as the v3 formatFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell./// SPEC §11); checked at pre-flight against formatFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell..env("LUMEN_PROJECT_ROOT", project_root)Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.out.extend_from_slice(b"RPGMV\0\0\0\0\x03\x01\x01\x01\x01\x01\x01");Fix: Put hosts, commands and strings in plain source. Decoding at runtime looks identical to malware and costs points on its own.<summary>Native binaries (Ghidra)</summary> Drives Ghidra 12 headless: one batch export per subject, parsed into evidence. Functions, names, strings, inventory Fix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.let app = fathom_mcp_server::http::router(state).merge(fathom_a2a::card::card_router(card));Fix: Namespace tool names and keep tool descriptions literal; no instructions to the agent inside them.// report a kill that never happened.Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.<summary>Tool classes</summary> | Class | Default | Rule | | --- | --- | --- | | `read` | allow | Static inspection; still confined to target roots | | `observeFix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.<summary>JavaScript and Electron</summary> oxc parses each module; parse errors are findings, not run failures. The module graph resolves relative specifiers; EFix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.let state = Arc::new(fathom_mcp_server::http::HttpState {Fix: Namespace tool names and keep tool descriptions literal; no instructions to the agent inside them.<summary>The evidence rule</summary> Every tool returns an evidence envelope: findings, each with a confidence class (`observed`, `derived`, `inferred`), the suFix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.assert!(url_has_credentials("https://user:[email protected]/x"));Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.format!("wait after kill failed: {e}"),Fix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.<summary>Engines are child processes</summary> Ghidra runs as `analyzeHeadless` batch jobs; nothing is reimplemented that an engine already does. Child processeFix: Remove the text that addresses the agent rather than the reader; describe behavior, do not instruct.parse_video_url("https://user:[email protected]/watch?v=s9f9oa3Ae6c").unwrap_err();Fix: Stop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.// Group-kill only a group we can prove we created: theFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.// `kill -KILL -<pid>` form once landed on the caller'sFix: Build commands from fixed arguments; do not interpolate untrusted strings into a shell.Cost to the agent
Tokens estimated at four characters each. Skills load their instruction text when triggered; the figure is what that costs.
Known vulnerabilities
Vulnerability advisories (OSV) cover npm and PyPI packages; this registry has no advisory database.
What would make this Verified
19 points to 80 · weights: behavior 45%, static 25%, prompt 15%, publisher 15%- Unclaimed publisher≈ +8 on the totalClaim qompassai: a verified publisher gets +50 on publisher trust. It is free and usually instant with GitHub sign-in.
- Hidden instructions≈ +6 on the totalRemove HTML comments, zero-width characters and white-on-white text that carry instructions. Everything the agent is meant to read should be visible to the person reading the file.
- Destructive file system command detected≈ +4 on the totalBuild commands from fixed arguments; do not interpolate untrusted strings into a shell.
- Access to sensitive file paths≈ +4 on the totalStop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.
- Hex-encoded string detected≈ +4 on the totalPut hosts, commands and strings in plain source. Decoding at runtime looks identical to malware and costs points on its own.
- URL with embedded credentials≈ +4 on the totalStop enumerating or reading credential locations; take secrets from a declared environment variable the user sets on purpose.