configs/default_policy.yaml

version: "1.0"
policy_name: "default"

default_action: ALLOW

ingress_rules:
  - name: block_prompt_injection
    description: "Detected prompt injection attempt"
    priority: 100
    action: DENY
    deny_message: "[LOBSTER TRAP] Blocked: prompt injection detected."
    conditions:
      - field: contains_injection_patterns
        match_type: boolean
        value: true

  - name: block_sensitive_paths
    description: "Prompt targets sensitive system paths"
    priority: 85
    action: DENY
    deny_message: "[LOBSTER TRAP] Blocked: sensitive path access denied."
    conditions:
      - field: contains_sensitive_paths
        match_type: boolean
        value: true

  - name: block_harm_violence
    description: "Prompt requests instructions for violence, weapons, or harmful substances"
    priority: 98
    action: DENY
    deny_message: "[LOBSTER TRAP] Blocked: request for harmful/violent content detected."
    conditions:
      - field: contains_harm_patterns
        match_type: boolean
        value: true

  - name: block_malware_request
    description: "Prompt requests creation of malware, exploits, or offensive tools"
    priority: 96
    action: DENY
    deny_message: "[LOBSTER TRAP] Blocked: malware/exploit generation request detected."
    conditions:
      - field: contains_malware_request
        match_type: boolean
        value: true

  - name: block_phishing_fraud
    description: "Prompt requests creation of phishing content or fraudulent materials"
    priority: 94
    action: DENY
    deny_message: "[LOBSTER TRAP] Blocked: phishing/fraud content request detected."
    conditions:
      - field: contains_phishing_patterns
        match_type: boolean
        value: true

  - name: block_data_exfiltration
    description: "Prompt contains data exfiltration patterns"
    priority: 92
    action: DENY
    deny_message: "[LOBSTER TRAP] Blocked: data exfiltration attempt detected."
    conditions:
      - field: contains_exfiltration
        match_type: boolean
        value: true

  - name: block_obfuscation_evasion
    description: "Prompt uses encoding or obfuscation to evade detection"
    priority: 90
    action: DENY
    deny_message: "[LOBSTER TRAP] Blocked: obfuscation/evasion technique detected."
    conditions:
      - field: contains_obfuscation
        match_type: boolean
        value: true

  - name: review_role_impersonation
    description: "Prompt attempts to assign a privileged role identity to the model"
    priority: 86
    action: HUMAN_REVIEW
    conditions:
      - field: contains_role_impersonation
        match_type: boolean
        value: true

  - name: block_pii_request
    description: "Prompt is requesting personal/sensitive information"
    priority: 82
    action: DENY
    deny_message: "[LOBSTER TRAP] Blocked: request for personal/sensitive information detected."
    conditions:
      - field: contains_pii_request
        match_type: boolean
        value: true

  - name: block_dangerous_commands
    description: "Dangerous system commands detected"
    priority: 80
    action: DENY
    deny_message: "[LOBSTER TRAP] Blocked: dangerous command detected."
    conditions:
      - field: contains_system_commands
        match_type: boolean
        value: true
      - field: risk_score
        match_type: threshold
        value: 0.3

  - name: review_high_risk
    description: "High risk score requires human review"
    priority: 70
    action: HUMAN_REVIEW
    conditions:
      - field: risk_score
        match_type: threshold
        value: 0.6

  - name: log_code_execution
    description: "Log code execution requests"
    priority: 30
    action: LOG
    conditions:
      - field: intent_category
        match_type: exact
        value: "code_execution"

egress_rules:
  - name: block_credential_leak
    description: "Model output contains credentials"
    priority: 100
    action: DENY
    deny_message: "[LOBSTER TRAP] Output blocked: contains credentials."
    conditions:
      - field: contains_credentials
        match_type: boolean
        value: true

  - name: block_pii_leak
    description: "Model output contains PII"
    priority: 90
    action: DENY
    deny_message: "[LOBSTER TRAP] Output blocked: contains PII."
    conditions:
      - field: contains_pii
        match_type: boolean
        value: true

rate_limits:
  requests_per_minute: 120
  requests_per_hour: 2000
  burst_threshold: 30

network:
  egress_policy: allowlist
  allowed_domains:
    - "api.openai.com"
    - "api.anthropic.com"
  denied_domains:
    - "*.onion"
    - "pastebin.com"

filesystem:
  denied_paths:
    - "/etc/**"
    - "/root/**"
    - "**/.ssh/**"
    - "**/.env"
    - "**/*secret*"
    - "**/*password*"
  allowed_read_paths:
    - "/home/*/documents/**"
    - "/tmp/agent_workspace/**"
  allowed_write_paths:
    - "/home/*/documents/agent_output/**"
    - "/tmp/agent_workspace/**"