io.github.Shrike-Security/shrike-mcp

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).

  tool "scan_response" description: names an outbound copy/exfil destination (BCC/exfil verb). tool "scan_command" description: names an outbound copy/exfil destination (BCC/exfil verb). tool "scan_web_search" description: names an outbound copy/exfil destination (BCC/exfil verb). tool "scan_a2a_message" description: names an outbound copy/exfil destination (BCC/exfil verb). tool "scan_agent_card" description: names an outbound copy/exfil destination (BCC/exfil verb). tool "scan_prompt" description: bare 'IMPORTANT:' prose in a description (tag form is high-confidence; prose is warn). tool "check_approval" description: bare 'IMPORTANT:' prose in a description (tag form is high-confidence; prose is warn). tool "reset_session" description: bare 'IMPORTANT:' prose in a description (tag form is high-confidence; prose is warn).

• Supply chain

  warn25% of grade

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

  install hook "prepare" present and not a recognized benign build command

• Credential hygiene

  fail15% of grade

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

  tool "scan_response" has a secret-shaped input parameter "pii_tokens" — secrets must never be tool parameters (the model would supply/handle the credential).

• Permission scope

  pass15% of grade

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

  No shell-exec, unbounded/sensitive filesystem, or download-and-run primitives found in scanned source. PASS = no static red flag. Static scanning is blind to code fetched or executed at runtime.

• Version behavior

  pass10% of grade

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

  vs prior 1.1.1 (major bump): new tool "scan_prompt" touches a sensitive surface (email/exec/net/fs); new tool "scan_response" touches a sensitive surface (email/exec/net/fs); new tool "scan_sql_query" touches a sensitive surface (email/exec/net/fs); new tool "scan_command" touches a sensitive surface (email/exec/net/fs); new tool "scan_file_write" touches a sensitive surface (email/exec/net/fs); new tool "scan_web_search" touches a sensitive surface (email/exec/net/fs); new tool "report_bypass" touches a sensitive surface (email/exec/net/fs); new tool "check_approval" appeared; new tool "reset_session" appeared; new tool "session_status" touches a sensitive surface (email/exec/net/fs); new tool "scan_a2a_message" touches a sensitive surface (email/exec/net/fs); new tool "scan_agent_card" touches a sensitive surface (email/exec/net/fs); new tool "scan_declare_scope" touches a sensitive surface (email/exec/net/fs); new tool "scan_mcp_schema" touches a sensitive surface (email/exec/net/fs). Low-severity drift (informational; consistent with the version bump).

• Transport config

  unscannable10% of grade

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

  local stdio server: no remote transport or auth surface, and no phone-home/telemetry signal found in source. Absence of transport findings is NOT a safety guarantee — phone-home hosts can be constructed at runtime and defeat static analysis.