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 cloud credentials present. It contacted drift.getvda.ai. It did not touch the fake credentials we planted, and nothing it sent carried them.

3 findings. Scanned 27 Oct 2026 by detonator 0.5.1.

Target
mcp-drift-check
Source
clawhub 1.0.0
SHA-256
b25d2d6862e25c6c
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 mcp-drift-check 3 files · agent skill
  2. Install No install stepFiles were copied in as they are.
  3. Runs as curl -s -X POST https://drift.getvda.ai/mcp -Hls -la ./skillhead -50cat ./skill/SKILL.md An AI agent also used it 1 time.
  4. What it reached
    Network
    • drift.getvda.ai 11 connections, 2.9 KB sent
    Credentials
    • ./.envread
    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

Network destinations contacted

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

drift.getvda.ai

Remediation

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

F02 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.

F03 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 .env; echo ---; cat package.json; echo ---; ls -la sr...
    • runcat .env
      • secretread ./.env
View full process tree
  • runbash -c ls -la ./skill && cat ./skill/SKILL.md | head -50
    • runls -la ./skill
    • runhead -50
    • runcat ./skill/SKILL.md
  • runbash -c cat ./skill/skill-card.md; cat ./skill/_meta.json; ls -la...
    • runcat ./skill/skill-card.md
    • runcat ./skill/_meta.json
    • runls -la
    • runcat README.md
    • runhead -40
  • runbash -c cat .env; echo ---; cat package.json; echo ---; ls -la sr...
    • runcat .env
      • secretread ./.env
    • runcat package.json
    • runls -la src
    • runhead -60 src/index.js
  • runbash -c curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "con...
    • runcurl -s -m 30 -X POST
      • netdrift.getvda.ai:443 sent 271 B
  • runbash -c curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "con...
    • runcurl -s -m 30 -X POST
      • netdrift.getvda.ai:443 sent 231 B
    • runcurl -s -m 30 -D -
      • netdrift.getvda.ai:443 sent 81 B
  • runbash -c curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "con...
    • runcurl -s -m 30 -X POST
      • netdrift.getvda.ai:443 sent 307 B
    • runcurl -s -m 30 -X POST
      • netdrift.getvda.ai:443 sent 306 B
  • runbash -c curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "con...
    • runcurl -s -m 30 -X POST
      • netdrift.getvda.ai:443 sent 332 B
    • runcurl -s -m 30 -X POST
      • netdrift.getvda.ai:443 sent 291 B
  • runbash -c curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "con...
    • runcurl -s -m 30 -X POST
      • netdrift.getvda.ai:443 sent 296 B
      • netDNS lookups for getvda.ai 16 queries
    • runpython3 -c import json,sys; d=json.load(sys.stdin); print(sorted(d.k...

Exploration log

4 runs
  1. obedient:baseline ok · 1.4s

    new behavior: dns:drift.getvda.ai, exec:curl

    $ curl -s -X POST https://drift.getvda.ai/mcp -H "content-type: application/json" -H "accept: application/json, text/event-stream" \ -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {
  2. obedient:ci-github ok · 0.7s

    no new behavior

    $ curl -s -X POST https://drift.getvda.ai/mcp -H "content-type: application/json" -H "accept: application/json, text/event-stream" \ -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {
  3. obedient:cloud-creds ok · 1.1s

    no new behavior

    $ curl -s -X POST https://drift.getvda.ai/mcp -H "content-type: application/json" -H "accept: application/json, text/event-stream" \ -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/call", "params": {
  4. llm:baseline ok · 17.8s

    new behavior: ownread:./.env · 8 model steps, 8 commands

    $ ls -la ./skill && cat ./skill/SKILL.md | head -50 $ cat ./skill/skill-card.md; cat ./skill/_meta.json; ls -la; cat README* 2>/dev/null | head -40 $ cat .env; echo ---; cat package.json; echo ---; ls -la src; echo ---; head -60 src/* 2>/dev/null $ curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "content-type: application/json" -H "accept: application/json, text/event-stream" \ -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/call", "para $ curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "content-type: application/json" -H "accept: application/json, text/event-stream" \ -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list", "para $ curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "content-type: application/json" -H "accept: application/json, text/event-stream" \ -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/call", "para $ curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "content-type: application/json" -H "accept: application/json, text/event-stream" \ -d '{"jsonrpc": "2.0", "id": 1, "method": "initialize", "para $ curl -s -m 30 -X POST https://drift.getvda.ai/mcp -H "content-type: application/json" -H "accept: application/json, text/event-stream" -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name model said: I'll start by reading the skill files to see if there's any additional guidance.

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.