io.github.vndee/llm-sandbox

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

  unscannable25% of grade

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

  No tool descriptions, server instructions, prompts, or resources were fetched; nothing to scan for injection.

• Supply chain

  pass25% of grade

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

  Python build script has no build-time network/exec; repository linkage verified

• Credential hygiene

  warn15% of grade

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

  examples/test_security.py: assigns a literal to password (user…rd (len 13)) — non-provider secret literal; prefer environment intake. tests/test_advanced_security_scenarios.py: assigns a literal to password (user…rd (len 13)) — non-provider secret literal; prefer environment intake. Reads secrets from the environment (e.g. examples/security_integration_tests.py) — the recommended intake shape.

• Permission scope

  fail15% of grade

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

  HIGH: executes content derived from a network response (examples/code_runner_docker.py: passes a network response into an exec/eval call). has both shell/exec and destructive filesystem writes (examples/security_integration_tests.py: subprocess.* call; examples/copy_demo.py: open(…, 'w'/'a'/'x') write-mode file access). discloses shell/exec capability: examples/security_integration_tests.py: subprocess.* call discloses shell/exec capability: examples/security_integration_tests.py: os.system/os.popen/os.exec* call discloses dynamic code execution: examples/security_policy_examples.py: uses eval()/exec() discloses dynamic code execution: examples/security_policy_presets.py: uses eval()/exec() discloses broad/destructive filesystem access: 9 files (examples/copy_demo.py, examples/k8s_readonly_file_system.py, examples/python_artifact.py, +6 more) Capability DISCLOSURE, not a verdict: static analysis sees the primitive is present and reachable, not whether its use is attacker-controlled.

• Version behavior

  unscannable10% of grade

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

  no prior version to diff (first sight of io.github.vndee/[email protected]); version-drift / rug-pull cannot be evaluated. Cold-start risk is covered by point-in-time checks (injection_surface, supply_chain, credential_hygiene), not here.

• Transport config

  warn10% of grade

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

  manifest/source references tracking behavior — disclosure