filters.json

{
  "label": "AI/Quarantine",
  "filters": [
    {
      "name": "Abuse-prone TLDs",
      "description": "Senders on TLDs with consistently high spam/phishing rates per Spamhaus and Google abuse data. Skip inbox and never mark important so they don't surface in default agent queries.",
      "query": "from:(*.skin OR *.top OR *.rest OR *.click OR *.cyou OR *.sbs OR *.monster OR *.zip OR *.mov)",
      "action": "quarantine",
      "remove_important": true
    },
    {
      "name": "Known prompt-injection phrases",
      "description": "Literal strings that have near-zero legitimate use in inbound email but appear frequently in jailbreak/injection content.",
      "query": "\"ignore previous instructions\" OR \"ignore all previous instructions\" OR \"ignore all instructions\" OR \"disregard previous instructions\" OR \"system prompt\" OR \"DAN mode\" OR \"jailbreak mode\" OR \"you are now a\" OR \"new instructions:\"",
      "action": "quarantine"
    },
    {
      "name": "Assistant name + action verb",
      "description": "Messages naming an assistant and issuing an action verb — a common shape for targeted prompt-injection attempts. Higher false-positive risk than the other filters; labeled only, not trashed.",
      "query": "(claude OR anthropic OR assistant OR gpt OR chatgpt) AND (execute OR curl OR \"email to\" OR \"forward to\" OR \"send to\")",
      "action": "quarantine"
    }
  ]
}