io.github.dshakes/distil

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

  fail25% of grade

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

  Python build script executes at build time: subprocess in build script

• Credential hygiene

  pass15% of grade

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

  Reads secrets from the environment (e.g. benchmarks/derc_live_compare.py) — the recommended intake shape. PASS = no static red flag; static analysis cannot prove the code honors env-based secret handling at runtime.

• 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 (benchmarks/fetch_real.py: subprocess.* call; benchmarks/fetch_real.py: outbound HTTP/socket call). has both shell/exec and destructive filesystem writes (benchmarks/fetch_real.py: subprocess.* call; benchmarks/long_horizon/run.py: open(…, 'w'/'a'/'x') write-mode file access). discloses shell/exec capability: 10 files (benchmarks/fetch_real.py, benchmarks/long_horizon/tools.py, benchmarks/swe_bench_e2e/preload_images.py, +7 more) discloses dynamic code execution: distil/codec/train_transformer.py: uses eval()/exec() discloses broad/destructive filesystem access: 7 files (benchmarks/long_horizon/run.py, benchmarks/swe_bench_e2e/compress_proxy.py, benchmarks/swe_bench_e2e/compress_proxy_openai.py, +4 more) 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 1.20.0 (minor bump): new capability primitive(s) in source: dynamic-eval, fs-write, http-client, net, subprocess; new hardcoded outbound host/BCC destination(s): 127.0.0.1, ai.google.dev, api.anthropic.com, api.openai.com, dshakes.github.io, evil.com, generativelanguage.googleapis.com, github.com, health.internal, img.shields.io, my-proxy.example.com, pypi.org, raw.githubusercontent.com, www.apple.com, www.ncei.noaa.gov, x.

• 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 telemetry behavior — disclosure