io.github.hermoso-ai/hermoso

What we checked

Six static checks, weighted by risk. Every result reflects only what could be observed in the published package and repository — never intent.

• Injection surface

  fail25% of grade

  Tool descriptions/manifest scanned for instruction-injection patterns (imperatives at the model, hidden text, 'ignore previous', data-exfil URLs).

  server instructions: instruction to hide activity from the user. tool "post_to_bluesky" param "captions" description: off-vendor URL in a description (…); capability text should not name where data goes. tool "post_to_meta" param "async" description: cross-tool redirection ('instead of call…'). tool "post_to_tiktok" description: imperative telling the model it must read/send/attach something. tool "draft_brand" description: bare 'IMPORTANT:' prose in a description (tag form is high-confidence; prose is warn). tool "competitor_teardown" description: imperative telling the model it must read/send/attach something. server instructions: imperative telling the model it must read/send/attach something.

• Supply chain

  pass25% of grade

  Package provenance: namespace verification, repo linkage, maintainer count, account age, postinstall scripts, typosquat distance.

  repository linkage verified

• Credential hygiene

  fail15% of grade

  How the server takes secrets (env vs plaintext config vs hardcoded); secrets appearing in tool schemas.

  tool "buy_credits" has a secret-shaped input parameter "quote_token" — secrets must never be tool parameters (the model would supply/handle the credential). Reads secrets from the environment (e.g. mcp/client.mjs) — the recommended intake shape.

• Permission scope

  warn15% of grade

  Declared tools vs. breadth (filesystem, network, exec); flags shell-exec and unbounded filesystem access.

  has both shell/exec and network egress — capable of download-and-run, though no such path was observed (bin/hermoso.mjs: imports child_process and calls exec/spawn; mcp/client.mjs: outbound HTTP/socket call). discloses shell/exec capability: bin/hermoso.mjs: imports child_process and calls exec/spawn Capability DISCLOSURE, not a verdict: static analysis sees the primitive is present and reachable, not whether its use is attacker-controlled.

• Version behavior

  fail10% of grade

  Diff of tool definitions between versions; new permissions or changed descriptions in a patch release (the postmark-mcp class).

  vs prior 0.1.193 (patch bump): tool "upload_file" description changed (wording only, no injection pattern); tool "post_to_meta" gained input property(ies): aiGenerated; tool "schedule_post" gained input property(ies): aiGenerated; tool "reschedule_post" gained input property(ies): aiGenerated; tool "post_to_youtube" gained input property(ies): aiGenerated; tool "post_to_tiktok" gained input property(ies): aiGenerated; new tool "find_tools" touches a sensitive surface (email/exec/net/fs); new tool "call_tool" touches a sensitive surface (email/exec/net/fs); new capability primitive(s) in source: child_process, http-client; new hardcoded outbound host/BCC destination(s): 127.0.0.1, app.hermoso.ai, github.com, hermoso.ai.

• Transport config

  warn10% of grade

  Remote servers: TLS and auth mode (none/token/OAuth). Local servers: whether the manifest indicates it phones home.

  sends telemetry to third-party SDK (google-analytics) — disclosure, verify opt-in; manifest/source references ANALYTICS behavior — disclosure