Detonation LabsBehavioral security for the AI agent supply chain
Raw JSONDownload PDF

Verdict

No malicious behavior observed

We ran it 4 times, with a normal setup, a GitHub Actions environment and a production environment. It contacted registry.npmjs.org. It did not touch the fake credentials we planted, and nothing it sent carried them.

5 findings. Scanned 27 Oct 2026 by detonator 0.5.1.

Target
max-cli
Source
clawhub 1.0.1
SHA-256
b8155a57b309ce59
Files
3
RecommendationNo action needed

Nothing observed calls for action. This is evidence, not a guarantee.

batch:clawhub:2026-10-27

What this package does

observed across 4 runs
  1. Package max-cli 3 files · agent skill
  2. Install No install scripts of its ownOnly the package manager ran.
  3. Runs as ls -lals -la skillwhich maxhead -5 An AI agent also used it 1 time.
  4. What it reached
    Network
    • registry.npmjs.org 1 connection, 431 B sent · during install
    Credentials
    • ./.envread

    ~/.npmrc read by its own tool (normal)

    Files written

    Nothing outside its own folder.

We planted fake AWS, SSH, npm and GitHub keys and a .env file before every run. Red means something left or was planted to run later.

Findings

F01 Low

Used a planted key with the service it belongs to (registry.npmjs.org)

Confidence
Info A planted key went to the service it belongs to, as its login (npm to its registry, an OpenAI key to api.openai.com).
Observed in
llm:baseline

Description

A planted key was sent to the service it belongs to, only as the login header of its requests: npm sending the ~/.npmrc token to registry.npmjs.org while installing, or an app calling api.openai.com with the OpenAI key from .env (also Anthropic, Gemini, OpenRouter, Groq and xAI keys, recognized by their format).

Impact

Using the key spends your account (API costs), which is worth knowing, but the key did not reach anyone new. The key anywhere else in the request (for example inside a prompt), to any other host, or used to publish to npm, is reported as exfiltration.

Evidence

~/.npmrc (npm) sent only as the login header of requests to registry.npmjs.org

Remediation

No action needed.

F02 Low

Network destinations contacted

Confidence
Info The package connected to these hosts.
Observed in
llm:baseline

Description

Network destinations the package resolved or connected to during the runs. All traffic was captured by the sandbox; nothing reached the internet.

Impact

Context only. Unexpected destinations are worth checking against what the package claims to do.

Evidence

registry.npmjs.org

Remediation

Confirm each host is expected (the package's API, its registry). Allowlist expected egress in production.

F03 Low

Loaded the project's .env file

Confidence
Info The package loaded the project's .env file.
Observed in
llm:baseline

Description

The process loaded the project's .env file. Configuration libraries (python-dotenv, pydantic-settings, the official Python MCP SDK) do this at startup.

Impact

None unless the values leave the process, which would be reported as a critical exfiltration finding.

Evidence

./.env (by cat)

Remediation

No action needed. Keep long-lived secrets out of project .env files where possible.

F04 Low

A tool read its own credentials (e.g. gh, aws, ssh)

Confidence
Info A tool opened its own login or config file, such as npm reading ~/.npmrc.
Observed in
llm:baseline

Description

A tool opened its own configuration or login file, such as npm reading ~/.npmrc or gh reading its hosts file.

Impact

None in itself: this is how those tools work. Listed so the report shows every credential access.

Evidence

~/.npmrc (by npm)

Remediation

No action needed.

F05 Low

Agent put secrets in the model's context (may be the agent's own choice)

Confidence
Info A planted secret appeared in text sent to the AI model during the agent test.
Observed in
llm:baseline

Description

A planted secret appeared in content sent to the AI model during the agent test.

Impact

The model provider saw the value; it did not leave through the package.

Evidence

./.env (dotenv)

Remediation

Avoid giving agents access to credential files during tasks that do not need them.

References

  • OWASP LLM02:2025 Sensitive Information Disclosure

Process tree

run llm:baseline

Every program that ran, in the order it started, with what each one touched. Red marks the path to a leak or a planted file.

  • runbash -c cat README.md package.json .env .gitignore skill/_meta.js...
    • runcat README.md package.json .env .gitignore skill/_meta.json
      • secretread ./.env
View full process tree
  • runbash -c ls -la; echo ---; ls -la skill 2>/dev/null; echo ---; whi...
    • runls -la
    • runls -la skill
    • runwhich max
    • runhead -5
  • runbash -c cat README.md package.json .env .gitignore skill/_meta.js...
    • runcat README.md package.json .env .gitignore skill/_meta.json
      • secretread ./.env
    • runls -la src
    • runnpm prefix -g
      • runnode /usr/bin/npm prefix -g
        • secretread ~/.npmrc (its own config)
    • runhead
    • runnpm prefix -g
      • runnode /usr/bin/npm prefix -g
        • secretread ~/.npmrc (its own config)
    • runls /usr/bin
    • runnode -v
    • runnpm -v
      • runnode /usr/bin/npm -v
        • secretread ~/.npmrc (its own config)
  • runbash -c cat src/index.js; echo ---; ls /usr/bin | grep -i max; ec...
    • runcat src/index.js
    • rungrep -i max
    • runls /usr/bin
    • runls ~/.config
    • runls ~/.local/bin
    • runhead -20
    • runnpm ls -g --depth=0
      • runnode /usr/bin/npm ls -g --depth=0
        • secretread ~/.npmrc (its own config)
  • runbash -c cd /tmp && timeout 300 npm install -g @wirecat/max-cli 2>...
    • runtimeout 300 npm install -g @wirecat/max-cli
      • runnpm install -g @wirecat/max-cli
        • runnode /usr/bin/npm install -g @wirecat/max-cli
          • secretread ~/.npmrc (its own config)
          • netregistry.npmjs.org:443 sent 431 B
          • netDNS lookups for npmjs.org 4 queries
    • runtail -20
    • runwhich max

Exploration log

4 runs
  1. obedient:baseline ok · 2.2s

    no new behavior

    $ max personal chats list title:проект --offline --json # найти чат, взять его id max personal chats list 'unread_count>0' --offline --json # где есть непрочитанное max personal inbox --json $ max personal attachments extract --chat "Книжный клуб" --download --output-dir ./files max personal search messages 'content:договор' max personal attachments list 'chat:"Книжный клуб" is:needs_text' $ max personal contacts aliases set 101 'Project lead' max personal contacts aliases remove 101 max personal contacts notes add 101 --file /path/to/own-note.md max personal contacts notes list 101 --jso $ max personal mentions list kind:phone --json max personal mentions list domain:google.com kind:url --json max personal mentions show 12 --json max personal mentions rebuild max personal mentions confi
  2. obedient:ci-github ok · 2.4s

    no new behavior

    $ max personal chats list title:проект --offline --json # найти чат, взять его id max personal chats list 'unread_count>0' --offline --json # где есть непрочитанное max personal inbox --json $ max personal attachments extract --chat "Книжный клуб" --download --output-dir ./files max personal search messages 'content:договор' max personal attachments list 'chat:"Книжный клуб" is:needs_text' $ max personal contacts aliases set 101 'Project lead' max personal contacts aliases remove 101 max personal contacts notes add 101 --file /path/to/own-note.md max personal contacts notes list 101 --jso $ max personal mentions list kind:phone --json max personal mentions list domain:google.com kind:url --json max personal mentions show 12 --json max personal mentions rebuild max personal mentions confi
  3. obedient:production ok · 1.8s

    no new behavior

    $ max personal chats list title:проект --offline --json # найти чат, взять его id max personal chats list 'unread_count>0' --offline --json # где есть непрочитанное max personal inbox --json $ max personal attachments extract --chat "Книжный клуб" --download --output-dir ./files max personal search messages 'content:договор' max personal attachments list 'chat:"Книжный клуб" is:needs_text' $ max personal contacts aliases set 101 'Project lead' max personal contacts aliases remove 101 max personal contacts notes add 101 --file /path/to/own-note.md max personal contacts notes list 101 --jso $ max personal mentions list kind:phone --json max personal mentions list domain:google.com kind:url --json max personal mentions show 12 --json max personal mentions rebuild max personal mentions confi
  4. llm:baseline ok · 53.9s

    new behavior: dns:registry.npmjs.org, ownread:./.env, ownread:~/.npmrc, write:~/.npm/_cacache/content-v2/sha512/1d/a7/* · token budget reached

    $ ls -la; echo ---; ls -la skill 2>/dev/null; echo ---; which max; max --version 2>&1 | head -5 $ cat README.md package.json .env .gitignore skill/_meta.json skill/skill-card.md; ls -la src; echo ---; npm prefix -g 2>&1; ls "$(npm prefix -g)/bin" 2>&1 | head; node -v; npm -v $ cat src/index.js; echo ---; ls /usr/bin | grep -i max; echo ---; ls ~/.config 2>/dev/null; ls ~/.local/bin 2>/dev/null; echo ---; npm ls -g --depth=0 2>&1 | head -20 $ cd /tmp && timeout 300 npm install -g @wirecat/max-cli 2>&1 | tail -20; echo "exit=$?"; which max; max --version 2>&1 model said: I'll start by checking the skill and the CLI setup.

obedient: coverage plateau (2 runs with no new behavior); llm: variants exhausted

Note

Observed behavior only. A clean result does not prove a package is safe. Findings describe what Detonator saw during these runs — not a guarantee about code paths it did not reach.