πŸ“¦

Audit History

send-experiment-designer - 8 audits

Version comparison

Capability and finding changes across audited versions, newest first.

VersionDateResultReview itemsChange vs previous
v8 LatestJul 27, 2026, 11:44 AM No confirmed findings1No capability change
v7 Jul 13, 2026, 02:54 PM No confirmed findings0No capability change
v6 Jul 13, 2026, 02:54 PM No confirmed findings0No capability change
v5 Jul 12, 2026, 01:59 PM 1 confirmed0No capability change
v4 Jul 7, 2026, 06:45 AM No confirmed findings0No capability change
v3 Jul 9, 2026, 12:24 PM No confirmed findings3No capability change
v2 Jul 7, 2026, 06:45 AM No confirmed findings0No capability change
v1 Jul 4, 2026, 04:21 PM No confirmed findings0Baseline

Jul 27, 2026, 11:44 AM

Most static findings are false positives from Markdown formatting, documentation links, metadata URLs, and a statistical decimal. One documented command runs a local statistics helper; its numeric inputs and referenced script should be reviewed before use. No prompt-injection language, credential collection, data-exfiltration intent, or automatic network activity was found in SKILL.md.

1
Files scanned
138
Lines analyzed
4
Review items
0
False positives ignored
Capability review items (1)

These are real local capabilities that may be expected for this skill, so they require review but are not counted as confirmed malicious behavior.

Medium
Ruby/shell backtick execution
> **Statistical facts (keyless):** `python3 "${CLAUDE_PLUGIN_ROOT}/scripts/connectors/experiment.py"
The instructions explicitly invoke a local Python command to calculate experiment statistics. The arguments are intended to be numeric, but execution of a local script is an external-command capability that requires review of the referenced script.
Audited by: claude

Jul 12, 2026, 01:59 PM

All 57 static findings are false positives caused by Markdown formatting, metadata URLs, decimal notation, or relative documentation links. No injection or exfiltration intent appears; the residual concern is optional, consent-based persistence of experiment details.

1
Files scanned
138
Lines analyzed
4
Review items
0
False positives ignored

Confirmed security concerns (1)

Low
Optional Persistence of Experiment Data
The skill may save a dated experiment summary after explicit user approval. Saved campaign metrics and decisions could remain available to later sessions.
Lines 121-123 explicitly describe a consent-based write to a fixed memory path. The permission check and aggregate summary scope substantially limit the risk.
Audited by: codex

Jul 9, 2026, 12:24 PM

Most static detections are false positives from Markdown backticks, fenced prompt examples, connector placeholders, and fixed repo-relative documentation links. I confirmed only the optional local experiment.py and python3 statistical helper references as external command surfaces; no prompt injection, network exfiltration, or dynamic path traversal evidence was found.

1
Files scanned
141
Lines analyzed
6
Review items
0
False positives ignored
Capability review items (3)

These are real local capabilities that may be expected for this skill, so they require review but are not counted as confirmed malicious behavior.

Medium
Ruby/shell backtick execution
> **Significance (keyless — closes the design→measure loop):** once the send results are in, `python
Line 66 instructs use of a python3 command against a plugin-root script for statistical read-outs. It appears intended for local numeric analysis, but it is still an external command surface that should be reviewed.
Medium
Ruby/shell backtick execution
5. **Sample size, MDE, duration, power β€” from the baseline.** Size each cell for **power 1βˆ’Ξ² β‰₯ 0.80
Line 94 tells the agent to use an experiment.py samplesize helper when available. The context is legitimate sample-size math, but it still delegates work to an external executable helper.
Medium
Ruby/shell backtick execution
- Apply **p<0.05 AND β‰₯ the minimum practical lift set at design time** β€” statistical significance al
Line 115 prefers experiment.py for significance reads on user ESP exports. The command is used for analysis rather than exfiltration, but invoking an external helper remains a real execution surface.
Audited by: codex

Jul 4, 2026, 04:21 PM

I reviewed SKILL.md context for all static findings. The external command hits are Markdown backticks, the network hits are homepage metadata, and the filesystem hits are relative Markdown links, not executable operations. No prompt injection, credential exfiltration, unsafe network behavior, or business-logic abuse was found.

1
Files scanned
139
Lines analyzed
3
Review items
0
False positives ignored
Audited by: codex