Incident Response Playbook for AI 安全 Breaches

# AI 安全 Incident Severity Levels
 
## SEV-1: Critical
**Definition**: Active 利用 causing immediate harm
**範例**:
- Model executing unauthorized financial transactions through tool abuse
- Bulk data exfiltration through systematic 提示詞注入
- Model generating content that creates immediate legal liability
- Production API key or credential leakage confirmed in the wild
 
**Response time**: Immediate (within 15 minutes)
**Who is notified**: CISO, Engineering Lead, Legal, On-call responders
**Containment**: Disable affected endpoint or switch to safe mode
 
## SEV-2: High
**Definition**: Confirmed 漏洞 under active 利用
**範例**:
- 系統提示詞 fully extracted and shared externally
- Sustained 提示詞注入 campaign bypassing 護欄
- Cross-user data leakage confirmed
- RAG pipeline serving poisoned content to users
 
**Response time**: Within 1 hour
**Who is notified**: 安全 team lead, Engineering lead
**Containment**: Enable enhanced filtering, rate limit affected users
 
## SEV-3: Medium
**Definition**: Confirmed 漏洞 with limited 利用
**範例**:
- Successful 越獄 producing policy-violating content
- Partial 系統提示詞 extraction
- Tool invocation with unauthorized parameters (no impact confirmed)
- Single-user data leakage incident
 
**Response time**: Within 4 hours
**Who is notified**: 安全 team, AI team
**Containment**: Add specific filter rules, monitor for escalation
 
## SEV-4: Low
**Definition**: Potential 漏洞 or minor policy violation
**範例**:
- Model producing mildly off-topic or inaccurate content
- Anomalous usage patterns that may indicate probing
- 安全 filter producing excessive false positives
- Minor configuration drift detected
 
**Response time**: Next business day
**Who is notified**: AI team
**Containment**: Monitor, add to backlog

# incident/triggers.py
"""Automated incident 偵測 triggers."""
from dataclasses import dataclass
from enum import Enum
from datetime import datetime
 
 
class TriggerSeverity(Enum):
    SEV1 = 1
    SEV2 = 2
    SEV3 = 3
    SEV4 = 4
 
 
@dataclass
class IncidentTrigger:
    name: str
    description: str
    severity: TriggerSeverity
    auto_contain: bool  # Whether to automatically trigger containment
    condition: str  # Human-readable condition
 
 
INCIDENT_TRIGGERS = [
    # SEV-1 Triggers
    IncidentTrigger(
        name="unauthorized_tool_execution",
        description="Model executed a tool with unauthorized parameters that resulted in a side effect",
        severity=TriggerSeverity.SEV1,
        auto_contain=True,
        condition="Tool execution log shows completed action with parameter validation bypass",
    ),
    IncidentTrigger(
        name="credential_leakage_confirmed",
        description="API key, 資料庫 credential, or secret detected in model response delivered to user",
        severity=TriggerSeverity.SEV1,
        auto_contain=True,
        condition="輸出 filter canary 符元 or credential pattern detected in delivered response",
    ),
    IncidentTrigger(
        name="bulk_data_extraction",
        description="Systematic extraction of data through repeated queries",
        severity=TriggerSeverity.SEV1,
        auto_contain=True,
        condition=">100 successful data queries from single user in 1 hour with unique results",
    ),
 
    # SEV-2 Triggers
    IncidentTrigger(
        name="sustained_injection_campaign",
        description="Coordinated 提示詞注入 attempts bypassing 護欄",
        severity=TriggerSeverity.SEV2,
        auto_contain=False,
        condition=">50 護欄 blocks from single user in 30 min with >10% bypass rate",
    ),
    IncidentTrigger(
        name="system_prompt_exposure",
        description="Full or substantial 系統提示詞 extracted by a user",
        severity=TriggerSeverity.SEV2,
        auto_contain=False,
        condition="Canary 符元 detected in 輸出 or >3 系統提示詞 fragments in single session",
    ),
    IncidentTrigger(
        name="cross_user_data_leakage",
        description="User A's data appearing in User B's responses",
        severity=TriggerSeverity.SEV2,
        auto_contain=True,
        condition="Response contains PII or identifiers belonging to a different user",
    ),
 
    # SEV-3 Triggers
    IncidentTrigger(
        name="jailbreak_success",
        description="Model produced content violating 安全 policy",
        severity=TriggerSeverity.SEV3,
        auto_contain=False,
        condition="輸出 content classifier flags delivered response as policy-violating",
    ),
    IncidentTrigger(
        name="partial_prompt_extraction",
        description="Fragments of 系統提示詞 detected in responses",
        severity=TriggerSeverity.SEV3,
        auto_contain=False,
        condition="1-2 系統提示詞 fragments detected in 輸出",
    ),
 
    # SEV-4 Triggers
    IncidentTrigger(
        name="anomalous_probing",
        description="Usage patterns consistent with 安全 probing",
        severity=TriggerSeverity.SEV4,
        auto_contain=False,
        condition="Anomaly detector flags user with >3 distinct probing patterns",
    ),
]

# incident/containment.py
"""Containment procedures for AI 安全 incidents."""
import logging
from dataclasses import dataclass
from enum import Enum
from datetime import datetime
 
logger = logging.getLogger(__name__)
 
 
class ContainmentAction(Enum):
    BLOCK_USER = "block_user"
    ENABLE_SAFE_MODE = "enable_safe_mode"
    DISABLE_TOOLS = "disable_tools"
    RATE_LIMIT = "rate_limit"
    DISABLE_ENDPOINT = "disable_endpoint"
    ROTATE_CREDENTIALS = "rotate_credentials"
    FLUSH_MEMORY = "flush_memory"
    RESTRICT_RAG = "restrict_rag"
 
 
@dataclass
class ContainmentPlan:
    severity: int
    actions: list[ContainmentAction]
    description: str
 
 
CONTAINMENT_PLANS = {
    1: ContainmentPlan(
        severity=1,
        actions=[
            ContainmentAction.BLOCK_USER,
            ContainmentAction.DISABLE_TOOLS,
            ContainmentAction.ROTATE_CREDENTIALS,
            ContainmentAction.ENABLE_SAFE_MODE,
        ],
        description=(
            "SEV-1: Block the offending user immediately. Disable 工具呼叫. "
            "Rotate any credentials that may have been exposed. Enable safe mode "
            "which returns only pre-approved responses."
        ),
    ),
    2: ContainmentPlan(
        severity=2,
        actions=[
            ContainmentAction.BLOCK_USER,
            ContainmentAction.RATE_LIMIT,
            ContainmentAction.FLUSH_MEMORY,
        ],
        description=(
            "SEV-2: Block the offending user. Apply aggressive rate limits to "
            "the affected endpoint. Flush conversation memory to prevent further "
            "context leakage."
        ),
    ),
    3: ContainmentPlan(
        severity=3,
        actions=[
            ContainmentAction.RATE_LIMIT,
        ],
        description=(
            "SEV-3: Apply rate limits to the offending user. Monitor for "
            "escalation. No immediate system-wide action needed."
        ),
    ),
    4: ContainmentPlan(
        severity=4,
        actions=[],
        description="SEV-4: Monitor and log. No immediate containment action.",
    ),
}
 
 
class ContainmentExecutor:
    """Execute containment actions."""
 
    def execute_plan(self, severity: int, user_id: str, incident_id: str):
        """Execute the containment plan for the given severity."""
        plan = CONTAINMENT_PLANS.get(severity)
        if not plan:
            logger.error(f"No containment plan for severity {severity}")
            return
 
        logger.critical(
            f"Executing containment plan for incident {incident_id} "
            f"(SEV-{severity}): {plan.description}"
        )
 
        for action in plan.actions:
            self._execute_action(action, user_id, incident_id)
 
    def _execute_action(self, action: ContainmentAction, user_id: str, incident_id: str):
        """Execute a single containment action."""
        logger.info(f"Executing {action.value} for user {user_id} (incident {incident_id})")
 
        if action == ContainmentAction.BLOCK_USER:
            # Add user to blocklist
            logger.warning(f"BLOCKED user {user_id}")
 
        elif action == ContainmentAction.DISABLE_TOOLS:
            # Disable 工具呼叫 system-wide
            logger.warning("DISABLED 工具呼叫 for all users")
 
        elif action == ContainmentAction.ROTATE_CREDENTIALS:
            # Trigger credential rotation
            logger.critical("CREDENTIAL ROTATION triggered -- manual action required")
 
        elif action == ContainmentAction.ENABLE_SAFE_MODE:
            # Switch to safe mode (pre-approved responses only)
            logger.warning("SAFE MODE enabled -- model responses restricted")
 
        elif action == ContainmentAction.FLUSH_MEMORY:
            # Clear conversation memory
            logger.warning(f"FLUSHED conversation memory for user {user_id}")
 
        elif action == ContainmentAction.RATE_LIMIT:
            # Apply aggressive rate limits
            logger.info(f"RATE LIMIT applied to user {user_id}: 1 req/min")

# incident/investigation.py
"""Investigation workflow for AI 安全 incidents."""
from dataclasses import dataclass, field
from datetime import datetime
 
 
@dataclass
class InvestigationChecklist:
    incident_id: str
    severity: int
    steps: list[dict] = field(default_factory=list)
 
    def add_step(self, step: str, status: str = "pending"):
        self.steps.append({
            "step": step,
            "status": status,
            "timestamp": datetime.now().isoformat(),
            "notes": "",
        })
 
 
def create_investigation_checklist(incident_id: str, severity: int,
                                    trigger_name: str) -> InvestigationChecklist:
    """Create an investigation checklist based on incident type."""
    checklist = InvestigationChecklist(incident_id=incident_id, severity=severity)
 
    # Common steps for all severities
    checklist.add_step("Verify the incident trigger is not a false positive")
    checklist.add_step("識別 the affected user(s) and session(s)")
    checklist.add_step("Collect conversation logs for affected sessions")
    checklist.add_step("Determine the attack technique used")
 
    if trigger_name in ("credential_leakage_confirmed", "system_prompt_exposure"):
        checklist.add_step("識別 exactly what information was exposed")
        checklist.add_step("Determine if the exposed information has been used elsewhere")
        checklist.add_step("Rotate all exposed credentials immediately")
        checklist.add_step("Search for the exposed data in public sources")
        checklist.add_step("評估 whether breach notification is required")
 
    if trigger_name in ("unauthorized_tool_execution", "bulk_data_extraction"):
        checklist.add_step("Review tool execution audit logs")
        checklist.add_step("識別 all actions taken by 模型 during the incident")
        checklist.add_step("評估 the impact of each unauthorized action")
        checklist.add_step("Determine if any actions need to be reversed")
        checklist.add_step("Verify tool 授權 controls")
 
    if trigger_name in ("cross_user_data_leakage",):
        checklist.add_step("識別 all affected users whose data was exposed")
        checklist.add_step("Determine the scope of data leakage")
        checklist.add_step("Check conversation memory for cross-contamination")
        checklist.add_step("Notify affected users per data breach policy")
 
    if trigger_name in ("sustained_injection_campaign",):
        checklist.add_step("Collect all attack payloads from the campaign")
        checklist.add_step("識別 which payloads successfully bypassed 防禦")
        checklist.add_step("Add successful payloads to the 護欄 測試 suite")
        checklist.add_step("Determine if 攻擊者 exfiltrated any data")
 
    # Common closing steps
    checklist.add_step("Document the root cause")
    checklist.add_step("實作 remediation")
    checklist.add_step("Verify remediation with 紅隊 retest")
    checklist.add_step("Update 護欄 rules based on findings")
    checklist.add_step("Schedule post-incident review")
 
    return checklist

# incident/validation.py
"""Remediation validation for AI 安全 incidents."""
import json
import requests
from dataclasses import dataclass
from datetime import datetime
 
 
@dataclass
class ValidationResult:
    test_name: str
    passed: bool
    details: str
 
 
class RemediationValidator:
    """Validate that incident remediation is effective."""
 
    def __init__(self, target_url: str, api_key: str):
        self.target_url = target_url
        self.api_key = api_key
 
    def validate_incident_remediation(
        self,
        incident_id: str,
        attack_payloads: list[str],
        success_indicators: list[str],
    ) -> list[ValidationResult]:
        """Re-run the original attack payloads and verify they are blocked."""
        results = []
 
        for i, payload in enumerate(attack_payloads):
            try:
                response = requests.post(
                    f"{self.target_url}/api/v1/chat",
                    json={"message": payload},
                    headers={"Authorization": f"Bearer {self.api_key}"},
                    timeout=30,
                )
 
                response_text = json.dumps(response.json())
                still_vulnerable = any(
                    indicator.lower() in response_text.lower()
                    for indicator in success_indicators
                )
 
                results.append(ValidationResult(
                    test_name=f"Payload {i+1} (incident {incident_id})",
                    passed=not still_vulnerable,
                    details=(
                        f"Payload: {payload[:100]}... | "
                        f"Status: {'FIXED' if not still_vulnerable else 'STILL VULNERABLE'}"
                    ),
                ))
 
            except Exception as e:
                results.append(ValidationResult(
                    test_name=f"Payload {i+1} (incident {incident_id})",
                    passed=False,
                    details=f"Error during validation: {e}",
                ))
 
        return results
 
    def generate_validation_report(self, results: list[ValidationResult]) -> str:
        """Generate a validation report."""
        passed = sum(1 for r in results if r.passed)
        total = len(results)
 
        report = f"""# Remediation Validation Report
**Date**: {datetime.now().strftime('%Y-%m-%d %H:%M')}
**Total tests**: {total}
**Passed**: {passed}
**Failed**: {total - passed}
**Status**: {'REMEDIATION VERIFIED' if passed == total else 'REMEDIATION INCOMPLETE'}
 
## 測試 Results
"""
        for result in results:
            status = "PASS" if result.passed else "FAIL"
            report += f"- [{status}] {result.test_name}: {result.details}\n"
 
        return report

# Post-Incident Review Template
 
## Incident 總結
- **Incident ID**: [ID]
- **Severity**: [SEV-1/2/3/4]
- **Duration**: [偵測 time to resolution time]
- **Trigger**: [What triggered the incident]
- **Impact**: [What was the actual impact]
 
## Timeline
| Time | Event | Actor |
|------|-------|-------|
| HH:MM | Incident detected by [trigger] | Automated |
| HH:MM | Containment initiated | [Responder name] |
| HH:MM | Investigation started | [Investigator name] |
| HH:MM | Root cause identified | [Investigator name] |
| HH:MM | Remediation deployed | [Engineer name] |
| HH:MM | Remediation validated | [Tester name] |
| HH:MM | Incident closed | [Lead name] |
 
## Root Cause Analysis
1. **What happened?** [Factual description]
2. **Why did it happen?** [Root cause -- not just the trigger]
3. **Why wasn't it detected sooner?** [監控 gaps]
4. **Why wasn't it prevented?** [防禦 gaps]
 
## What Went Well
- [List things that worked during the response]
 
## What Needs Improvement
- [List things that should be improved]
 
## Action Items
| Action | Owner | Due Date | Status |
|--------|-------|----------|--------|
| Add attack payload to regression 測試 suite | [Name] | [Date] | Pending |
| Update 護欄 rules for this attack pattern | [Name] | [Date] | Pending |
| Improve 監控 to detect this pattern earlier | [Name] | [Date] | Pending |
| Review and update incident response playbook | [Name] | [Date] | Pending |
| Schedule follow-up 紅隊 測試 | [Name] | [Date] | Pending |