Local-First LLM

· v1.0.0

Medium Risk

"Routes LLM requests to a local model (Ollama, LM Studio, llamafile) before falling back to cloud APIs. Tracks token savings and cost avoidance in a persistent dashboard. Use when: (1) user asks to run a task with a local model first, (2) user wants to reduce cloud API costs or keep requests private, (3) user asks to see their token savings or LLM routing dashboard, (4) any request where local-vs-cloud routing should be decided automatically. Supports Ollama, LM Studio, and llamafile as local providers."

H:3 D:4 A:1 C:1

⚠️ Hazard Flags

EXEC FS_READ_WORKSPACE FS_WRITE_WORKSPACE FS_WRITE_USER NET_EGRESS_ANY PI_WEB

📋 Capabilities

Execution

✅ Shell execution
❌ Code execution
❌ Install dependencies
❌ Persistence
Privilege: user

Filesystem

✅ Read workspace
✅ Write workspace
❌ Read home
✅ Write home
❌ Read system
❌ Delete

Network

Egress: any
❌ Ingress

Credentials

❌ Environment vars
❌ Credential files
❌ Browser data
❌ Keychain

Actions

❌ send messages❌ post public❌ purchase❌ transfer money❌ deploy❌ delete external

🔒 Containment

Level: maximum

Required:

SANDBOX_CONTAINER: Code execution capability

Recommended:

LOG_ACTIONS: Audit trail for all actions

⚡ Risks

Social engineering indicators: SOCIAL_ENG_VAGUE_DESCRIPTION low

Mitigation: Provide clear, detailed description of skill functionality

Tool poisoning: hidden behaviors detected (MCP_TOOL_POISONING_SENSITIVE_DATA, MCP_TOOL_POISONING_REMOTE_STORAGE) high

Mitigation: Remove references to sensitive data collection.

Data exfiltration patterns: DATA_EXFIL_NETWORK_REQUESTS medium

Mitigation: Ensure network access is necessary and documented

Want a deeper analysis?

This report was generated by static analysis. Get an LLM-powered deep review with behavioral reasoning and attack surface mapping.

🧠 Deep Analysis — $5.00

🚨 Incident Response

Kill switch: Stop the agent process

Containment: Review logs for unexpected actions

Recovery: Depends on skill capabilities

📄 Raw SSDS JSON click to expand

{
  "meta": {
    "document_id": "ssds:auto:local-first-llm:1.0.0",
    "ssds_version": "0.2.0",
    "scanner_version": "0.4.0+fe6fd9123d50",
    "created_at": "2026-03-05T15:36:17.332Z",
    "created_by": {
      "agent": "safeagentskills-cli/generate-ssds"
    },
    "language": "en",
    "notes": "Auto-generated SSDS. Manual review recommended."
  },
  "skill": {
    "name": "Local-First LLM",
    "version": "1.0.0",
    "format": "agent_skill",
    "description": "\"Routes LLM requests to a local model (Ollama, LM Studio, llamafile) before falling back to cloud APIs. Tracks token savings and cost avoidance in a persistent dashboard. Use when: (1) user asks to run a task with a local model first, (2) user wants to reduce cloud API costs or keep requests private, (3) user asks to see their token savings or LLM routing dashboard, (4) any request where local-vs-cloud routing should be decided automatically. Supports Ollama, LM Studio, and llamafile as local providers.\"",
    "publisher": "unknown",
    "source": {
      "channel": "local"
    },
    "artifact": {
      "sha256": "60d943ae6734ec502f10932086edf476d0e953a5529585de1468337e2d7a7f78",
      "hash_method": "files_sorted"
    }
  },
  "capabilities": {
    "execution": {
      "can_exec_shell": true,
      "can_exec_code": false,
      "privilege_level": "user",
      "can_install_deps": false,
      "can_persist": false
    },
    "filesystem": {
      "reads_workspace": true,
      "reads_user_home": false,
      "reads_system": false,
      "writes_workspace": true,
      "writes_user_home": true,
      "writes_system": false,
      "can_delete": false
    },
    "network": {
      "egress": "any",
      "ingress": false
    },
    "credentials": {
      "reads_env_vars": false,
      "reads_credential_files": false,
      "reads_browser_data": false,
      "reads_keychain": false
    },
    "services": [],
    "actions": {
      "can_send_messages": false,
      "can_post_public": false,
      "can_purchase": false,
      "can_transfer_money": false,
      "can_deploy": false,
      "can_delete_external": false
    },
    "prompt_injection_surfaces": [
      "web"
    ],
    "content_types": [
      "general"
    ]
  },
  "hazards": {
    "hdac": {
      "H": 3,
      "D": 4,
      "A": 1,
      "C": 1
    },
    "flags": [
      "EXEC",
      "FS_READ_WORKSPACE",
      "FS_WRITE_WORKSPACE",
      "FS_WRITE_USER",
      "NET_EGRESS_ANY",
      "PI_WEB"
    ],
    "custom_flags": [
      {
        "code": "SOCIAL_ENGINEERING",
        "name": "Social Engineering Risk",
        "description": "SOCIAL_ENG_VAGUE_DESCRIPTION: Skill description is too vague or missing"
      },
      {
        "code": "TOOL_POISONING",
        "name": "Tool Poisoning",
        "description": "Hidden secondary behavior detected: MCP_TOOL_POISONING_SENSITIVE_DATA, MCP_TOOL_POISONING_REMOTE_STORAGE"
      },
      {
        "code": "DATA_EXFILTRATION",
        "name": "Data Exfiltration Risk",
        "description": "DATA_EXFIL_NETWORK_REQUESTS: HTTP client library imports that enable external communication"
      }
    ],
    "confidence": {
      "level": "medium",
      "basis": [
        "static_analysis"
      ],
      "notes": "Detected 3 security patterns (4 vendored rule hits). Review recommended."
    },
    "rationale": {
      "H": "H3: Shell/code execution or persistence detected",
      "D": "D4: Critical: Credential theft or data exfiltration",
      "A": "A1: Local side effects only",
      "C": "C1: General content"
    }
  },
  "containment": {
    "level": "maximum",
    "required": [
      {
        "control": "SANDBOX_CONTAINER",
        "reason": "Code execution capability"
      }
    ],
    "recommended": [
      {
        "control": "LOG_ACTIONS",
        "reason": "Audit trail for all actions"
      }
    ],
    "uncontained_risk": "Risk level depends on manual review of actual capabilities."
  },
  "risks": {
    "risks": [
      {
        "risk": "Social engineering indicators: SOCIAL_ENG_VAGUE_DESCRIPTION",
        "severity": "low",
        "mitigation": "Provide clear, detailed description of skill functionality"
      },
      {
        "risk": "Tool poisoning: hidden behaviors detected (MCP_TOOL_POISONING_SENSITIVE_DATA, MCP_TOOL_POISONING_REMOTE_STORAGE)",
        "severity": "high",
        "mitigation": "Remove references to sensitive data collection."
      },
      {
        "risk": "Data exfiltration patterns: DATA_EXFIL_NETWORK_REQUESTS",
        "severity": "medium",
        "mitigation": "Ensure network access is necessary and documented"
      }
    ],
    "limitations": [
      "Static analysis only - runtime behavior not verified"
    ]
  },
  "incident_response": {
    "kill_switch": [
      "Stop the agent process"
    ],
    "containment": [
      "Review logs for unexpected actions"
    ],
    "recovery": [
      "Depends on skill capabilities"
    ]
  },
  "evidence": [
    {
      "evidence_id": "EV:file-1",
      "type": "file_excerpt",
      "title": "scripts/check_local.py",
      "file_path": "scripts/check_local.py"
    },
    {
      "evidence_id": "EV:file-2",
      "type": "file_excerpt",
      "title": "scripts/dashboard.py",
      "file_path": "scripts/dashboard.py"
    },
    {
      "evidence_id": "EV:file-3",
      "type": "file_excerpt",
      "title": "scripts/route_request.py",
      "file_path": "scripts/route_request.py"
    },
    {
      "evidence_id": "EV:file-4",
      "type": "file_excerpt",
      "title": "scripts/track_savings.py",
      "file_path": "scripts/track_savings.py"
    },
    {
      "evidence_id": "EV:file-5",
      "type": "file_excerpt",
      "title": "SKILL.md",
      "file_path": "SKILL.md"
    },
    {
      "evidence_id": "EV:file-6",
      "type": "file_excerpt",
      "title": "_meta.json",
      "file_path": "_meta.json"
    },
    {
      "evidence_id": "EV:cisco-1",
      "type": "file_excerpt",
      "title": "DATA_EXFIL_NETWORK_REQUESTS [MEDIUM] scripts/check_local.py:8: import urllib.request",
      "file_path": "scripts/check_local.py"
    },
    {
      "evidence_id": "EV:cisco-2",
      "type": "file_excerpt",
      "title": "SOCIAL_ENG_VAGUE_DESCRIPTION [LOW] SKILL.md:1: ---",
      "file_path": "SKILL.md"
    },
    {
      "evidence_id": "EV:cisco-3",
      "type": "file_excerpt",
      "title": "MCP_TOOL_POISONING_SENSITIVE_DATA [HIGH] SKILL.md:3: description: \"Routes LLM requests to a local model (Ollama, LM Studio, llamafile",
      "file_path": "SKILL.md"
    },
    {
      "evidence_id": "EV:cisco-4",
      "type": "file_excerpt",
      "title": "MCP_TOOL_POISONING_REMOTE_STORAGE [MEDIUM] SKILL.md:78: | Prompt contains sensitive data (`password`, `secret`, `api key`, `ssn`, etc.) ",
      "file_path": "SKILL.md"
    }
  ]
}