中文 EN

扫描报告

扫描新文件

✓ 可信 — 风险评分 5/100
上次扫描：1 天前 重新扫描
5 /100
lobster-continuous-learning-v2
Instinct-based learning system that observes sessions, creates atomic instincts with confidence scoring, and evolves them into skills/commands/agents
Legitimate continuous learning system with no malicious behavior. Observations stay local, secret scrubbing is implemented, and all shell/network operations are documented and necessary for the feature.
技能名称lobster-continuous-learning-v2
分析耗时64.3s
引擎pi
✓
可以安装
No action required. Skill is safe to use.

安全发现 2 项

严重性 安全发现 位置
提示
Secret Scrubbing Regex May Miss Edge Cases 敏感访问
The _SECRET_RE pattern in observe.sh may not catch all credential formats (e.g., Bearer tokens without proper spacing). However, observations stay local and no exfiltration occurs.
r"(?i)(api[_-]?key|token|secret|password|authorization|credentials?|auth)"
→ Consider expanding regex to catch more patterns like 'Bearer ', 'ghp_', 'sk-' prefixes
 hooks/observe.sh:198
提示
External URL Reference 文档欺骗
SKILL.md contains a reference to an external Twitter/X URL (https://x.com/affaanmustafa/status/...). This is a documentation link, not executable code.
https://x.com/affaanmustafa/status/2014040193557471352
→ No action needed - this is just a documentation link
 SKILL.md:361
资源类型声明权限推断权限状态证据
文件系统 READ WRITE ✓ 一致 Write operations limited to ~/.claude/homunculus/ project storage directories
命令执行 READ READ ✓ 一致 subprocess.run for git detection only, no arbitrary command execution
网络访问 READ READ ✓ 一致 URL fetch only for instinct import, documented in SKILL.md
环境变量 NONE READ ✓ 一致 Reads CLAUDE_PROJECT_DIR and git-related vars for project detection
2 项发现
🔗
中危 外部 URL 外部 URL
https://x.com/affaanmustafa/status/2014040193557471352
SKILL.md:361
📧
提示 邮箱 邮箱地址
[email protected]
scripts/detect-project.sh:96

目录结构

10 文件 · 154.4 KB · 4302 行
Python 2f · 2410L Shell 5f · 1321L Markdown 2f · 563L JSON 1f · 8L
├─ ▾ 📁 agents
│ ├─ 🔧 observer-loop.sh Shell 271L · 9.2 KB
│ ├─ 📝 observer.md Markdown 198L · 7.2 KB
│ ├─ 🔧 session-guardian.sh Shell 150L · 6.2 KB
│ └─ 🔧 start-observer.sh Shell 244L · 7.3 KB
├─ ▾ 📁 hooks
│ └─ 🔧 observe.sh Shell 428L · 15.8 KB
├─ ▾ 📁 scripts
│ ├─ 🔧 detect-project.sh Shell 228L · 7.6 KB
│ ├─ 🐍 instinct-cli.py Python 1426L · 56.4 KB
│ └─ 🐍 test_parse_instinct.py Python 984L · 32.3 KB
├─ 🔑 config.json ⚠ JSON 8L · 135 B
└─ 📝 SKILL.md Markdown 365L · 12.3 KB

依赖分析 1 项

包名版本来源已知漏洞备注
requests N/A pip 否 No external dependencies used - only stdlib

安全亮点

✓ Secret scrubbing implemented to redact API keys, tokens, passwords before persisting observations
✓ Git credential stripping in detect-project.sh removes embedded credentials from remote URLs
✓ Path traversal protection blocks access to system directories (/etc, /usr, /bin, etc.)
✓ Project-scoped storage isolates data between different projects
✓ Observer agent uses minimal toolset (Read, Write only) with controlled Haiku model
✓ PID validation prevents signaling invalid process IDs
✓ Atomic file writes with tempfile + os.replace prevent corruption
✓ File locking (fcntl) prevents concurrent write race conditions
✓ No base64-encoded payloads or obfuscation techniques
✓ No ~/.ssh, ~/.aws, or other sensitive path access
✓ No curl|bash or remote script execution patterns
✓ No C2 or data exfiltration - all observations stay local
✓ Session guardian prevents observer from running during automated sessions