Compétences self-improving-agent Historique des audits
📦

Historique des audits

self-improving-agent - 2 audits

Version de l’audit 2

Dernier Risque élevé

Jun 28, 2026, 09:47 PM

Static analysis produced many alerts, but most blocker-level items are false positives from markdown examples, diagrams, and research links. The confirmed risk is high because this community skill requests broad file-editing and Bash capabilities, teaches self-modification of skills, and documents hooks that can log tool input and command output.

10
Fichiers analysés
1,037
Lignes analysées
9
résultats
codex
Audité par

Problèmes à risque élevé (2)

SKILL.md:3-4SKILL.md:33-39SKILL.md:231-269
Broad Self-Modifying Agent Permissions
The community skill requests Bash, Write, Edit, and WebSearch access while instructing the agent to update skill files, memory files, and related guidance after skill events. This is powerful automation, but it can change repository behavior beyond the immediate user task.
README.md:54-65SKILL.md:323-356
Hook Configuration Can Log Sensitive Tool Data
The hook examples pass tool input and Bash output into local shell commands for logging. Tool input and output can include secrets, paths, or proprietary data, so this pattern needs explicit user approval and log handling controls.
Problèmes à risque moyen (1)
README.md:22-27README.md:41-42SKILL.md:321-365
Home Directory Configuration Changes
The documentation instructs users to create a symlink under ~/.claude and edit Claude Code settings. This is a legitimate installation pattern, but it changes global agent behavior for the user account.
Problèmes à risque faible (3)
README.md:41-42README.md:54-65SKILL.md:325-363
Shell Command Findings Are Mostly Documentation
Most external command detections are markdown examples, diagrams, or hook configuration snippets. I did not find unquoted command construction or instructions to execute untrusted input through a shell.
SKILL.md:47-50SKILL.md:404-408references/appendix.md:127-131
Network Findings Are Research Links, Not Exfiltration
The hardcoded URLs point to research references and articles. I did not find evidence that the skill sends local data to those URLs.
SKILL.md:54-76SKILL.md:404-408references/appendix.md:127-131memory/semantic-patterns.json:253-270
Weak Crypto, C2, and Credential Alerts Appear False Positive
The weak cryptography, C2 keyword, Windows SAM, and reconnaissance alerts match prose, diagrams, markdown paths, or article titles. No evidence found of credential theft, cryptographic downgrade, C2 behavior, or system reconnaissance logic.

Facteurs de risque

⚙️ Commandes externes (3)
README.md:41-42 README.md:54-65 SKILL.md:323-356
🌐 Accès réseau (3)
SKILL.md:47-50 SKILL.md:404-408 references/appendix.md:127-131
📁 Accès au système de fichiers (3)
README.md:22-27 README.md:41-42 SKILL.md:321-365

Motifs détectés

Broad Self-Modifying Agent PermissionsHook Configuration Can Log Sensitive Tool Data

Version de l’audit 1

Sûr

Mar 17, 2026, 08:23 AM

This is a legitimate Claude Code skill for self-improvement. The static scanner flagged 121 potential issues, but after semantic evaluation, all are false positives. External command detections are markdown backticks in documentation. Network detections are academic reference URLs. Filesystem detections are standard Claude Code config paths. No actual security risks present.

10
Fichiers analysés
1,037
Lignes analysées
0
résultats
claude
Audité par
Aucun problème de sécurité trouvé
← Retour à la compétence