feat: 完成 issue #48 DLP 敏感数据拦截(目标 100% 拦截)

This commit is contained in:
2026-08-04 17:15:02 +08:00
parent 4c3a9fdbe6
commit fa5523a37d
6 changed files with 838 additions and 0 deletions
+76
View File
@@ -0,0 +1,76 @@
# iAOP-Core · LLM 网关(LLM Gateway)
对应 PRD 5.4「④ LLM 网关 + RAG」与 EPIC #6(Issue #48 等子任务):
本地 70B(敏感/核心)+ 云端 API(脱敏/通用)**混合**,安全分级路由,
**敏感数据本地闭环,仅脱敏/公开内容可走云端 API**(数据不出厂)。
## 模块结构
```
core/llm-gateway/
├── __init__.py 包入口(导出 DLP 引擎 API)
├── dlp.py DLP 敏感数据拦截引擎(Issue #48)
├── config/
│ └── dlp.template.yaml 模板 DLP 规则资产(ti-cl4 示例,换行业只改它)
└── tests/
├── _bootstrap.py 测试引导(目录含连字符,挂载包名 llm_gateway)
└── test_dlp.py DLP 引擎单元测试
```
## DLP 敏感数据拦截(Issue #48)
出站防线:任何要发往**云端**的内容(用户 query / RAG 检索到的 context /
模型输出)先经 `DlpEngine.check_outbound()` 检查,**命中任一 block 规则
即整包拦截**(fail-closed),并落结构化审计日志(全量)。
```python
from llm_gateway.dlp import DlpEngine
engine = DlpEngine.from_template_config("config/dlp.template.yaml")
result = engine.check_outbound({
"query": user_query, # 用户提问
"context": rag_context, # RAG 命中文档片段
"output": llm_answer, # 模型生成结果
})
if result.blocked:
# 不发往云端:转本地处理 / 转人工确认 / 拒绝(PRD 5.4 异常时转人工)
engine.drain_audit() # 取走审计记录(对接外部审计管道)
```
### 语义(目标 DLP 拦截率 100%)
- **命中即拦截**:`decision="block"`、`reason="hit"`;
- **未配置即拦截**:引擎无任何规则(含内置保底)且 `fail_closed=True` 时
`reason="no_rules"`——未配置 = 不安全,保守拒绝出站;
- **审计不落明文敏感内容**:命中明细只记录规则名 / 类别 / 所属 part /
位置与 `<category>` 脱敏占位。
### 规则来源
1. **内置保底**(`DLP_DEFAULT_RULES`,内核自带):通用 PII 与凭证
(身份证、手机号、邮箱、IPv4、云访问密钥 AK、密钥键值对)+ 示例工艺词;
2. **模板资产**(`config/dlp.template.yaml`):行业敏感资产(工艺参数 /
配方 / 关键设备 / 人员岗位),同名规则覆盖内置实现模板定制。
规则两种匹配方式:`keyword`(大小写不敏感子串)与 `regex`(正则)。
### 验收对照(PRD 5.4 / 父 Issue #6)
| 验收项 | 实现 |
|--------|------|
| DLP 拦截率 100% | fail-closed:命中即 block;未配置规则也 block |
| 敏感数据本地闭环 | 出站(云端通道)必经 check_outbound 检查 |
| 敏感词/DLP 规则为配置点 | `config/dlp.template.yaml`,换行业只改资产 |
| 审计日志全量(NFR 安全-数据) | 每次检查落一条结构化审计,可 drain 对接管道 |
## 运行测试
```bash
cd core/llm-gateway
python -m unittest discover -s tests -v
```
## 后续子任务(EPIC #6 拆分,待扩展)
- 敏感度路由(准确率 ≥ 96.5%)与路由准确率评估脚本(Issue #49);
- Prompt 版本管理与幻觉/事实性校验中间件(Issue #47)。
+34
View File
@@ -0,0 +1,34 @@
# -*- coding: utf-8 -*-
"""iAOP-Core · LLM 网关(LLM Gateway)—— 混合 LLM 的安全出站防线。
对应 PRD 5.4「④ LLM 网关 + RAG」与 EPIC #6:
本地 70B(敏感/核心)+ 云端 API(脱敏/通用)混合,敏感数据**本地闭环**。
当前子模块(Issue #48):
- dlp DLP 敏感数据拦截引擎:出站内容(query / RAG context / 模型输出)
发往云端前做敏感规则检查,命中即拦截(目标 100% 拦截),全量审计。
后续子任务(EPIC #6 拆分,将在本包扩展):
- 敏感度路由(router)、Prompt 版本管理与幻觉校验中间件、路由准确率评估。
测试:`python -m unittest discover -s tests -v`(在 core/llm-gateway 目录下执行)。
"""
__version__ = "0.1.0"
from .dlp import (
DLP_DEFAULT_RULES,
DlpEngine,
DlpHit,
DlpResult,
DlpRule,
DlpRuleKind,
)
__all__ = [
"DlpRuleKind",
"DlpRule",
"DlpHit",
"DlpResult",
"DlpEngine",
"DLP_DEFAULT_RULES",
]
+51
View File
@@ -0,0 +1,51 @@
# -*- coding: utf-8 -*-
# 模板 DLP 规则资产示例:ti-cl4(氯化车间/海绵钛,Template-Ti 一期)。
#
# 说明:
# - 这是「敏感词/DLP 规则」配置点(PRD 5.4):换行业只改本文件,内核零改动;
# - mode: fail-closed —— 命中任一规则即拦截出站(目标 DLP 拦截率 100%);
# - kind: keyword 大小写不敏感子串匹配;regex 正则匹配;
# - 通用 PII/凭证规则由内核内置保底(DLP_DEFAULT_RULES),无需重复配置;
# 本文件只补充**行业敏感资产**(工艺参数 / 配方 / 关键设备 / 人员等)。
template: ti-cl4
version: 1.0.0
mode: fail-closed
rules:
# ---- 工艺敏感参数(Ti-1 氯化车间) ----
- name: proc_cl2_flow
category: process-parameter
kind: keyword
pattern: 氯气流量
description: 氯气流量参数(工艺敏感)
- name: proc_furnace_temp
category: process-parameter
kind: keyword
pattern: 炉温
description: 炉温参数(工艺敏感)
- name: proc_feeding_ratio
category: process-parameter
kind: keyword
pattern: 加料比
description: 加料配比参数(工艺敏感)
- name: proc_ti_purity
category: process-parameter
kind: keyword
pattern: 钛纯度
description: 产品质量指标(钛纯度)
- name: proc_formula_detail
category: process-parameter
kind: regex
pattern: '配比\s*[::]?\s*[\d.]+%?'
description: 配方配比数值(结构化敏感数据)
# ---- 设备 / 工艺关键点(内部敏感命名) ----
- name: proc_furnace_id
category: process-asset
kind: regex
pattern: '沸腾氯化炉\s*[A-Z]?\d+'
description: 关键设备编号(沸腾氯化炉)
# ---- 人员 / 排班(内部敏感) ----
- name: staff_contact
category: personnel
kind: keyword
pattern: 值班长
description: 岗位/人员信息(内部敏感)
+446
View File
@@ -0,0 +1,446 @@
# -*- coding: utf-8 -*-
"""iAOP-Core · LLM 网关 —— DLP 敏感数据拦截(Issue #48,EPIC #6 子任务)。
对应 PRD 5.4「④ LLM 网关 + RAG」:
- 数据不出厂:敏感数据**本地闭环**,仅脱敏/公开内容可走云端 API;
- 验收:**DLP 拦截率 100%**,敏感词/DLP 规则作为模板配置点;
- 用户操作流程:用户提问 → 路由判断敏感级 → 本地/云端生成 →
RAG 溯源校验 → 返回答案(出站前经 DLP 检查)。
本模块实现 **出站 DLP 防线**:任何要发往**云端**的内容
(用户 query / RAG 检索到的 context / 模型输出)先经 `DlpEngine.check_outbound`
检查,命中任一 block 规则即**整包拦截**(fail-closed,目标 100% 拦截),
并落**结构化审计日志**(全量,PRD NFR 安全-数据)。
规则来源:
- 模板资产 `config/dlp.template.yaml`(行业敏感词 + 通用 PII 正则),
换行业只改该资产,内核零改动(对齐 rag-kb 模板化思想);
- `DLP_DEFAULT_RULES` 内置保底规则:即使未加载任何配置,
通用 PII(身份证/手机号/邮箱/IP/访问密钥等)也默认生效。
测试:`python -m unittest discover -s tests -v`(在 core/llm-gateway 目录下执行)。
"""
from __future__ import annotations
import re
import uuid
from dataclasses import dataclass, field
from datetime import datetime, timezone
from typing import Callable, Dict, List, Optional, Tuple
# ---------------------------------------------------------------------------
# 轻量 YAML 子集解析(零第三方依赖,递归下降):map / list / scalar / 注释。
# 与 rag-kb/templating.py 同款(模块内自持一份,保持模块零耦合)。
# 足以解析 `config/dlp.template.yaml` 这类模板资产。
# ---------------------------------------------------------------------------
def _parse_scalar(text: str) -> str:
"""去掉标量两侧引号与行内注释(`key: value # comment`)。"""
t = text.split(" #", 1)[0].strip()
if len(t) >= 2 and t[0] == t[-1] and t[0] in ("'", '"'):
return t[1:-1]
return t
def _strip_comments(lines: List[str]) -> List[Tuple[str, int]]:
"""剔除空行与整行注释,保留行号(1 起)用于报错定位。"""
out = []
for i, ln in enumerate(lines):
s = ln.strip()
if not s or s.startswith("#"):
continue
out.append((ln, i + 1))
return out
def _parse_node(lines: List[Tuple[str, int]], i: int, indent: int):
"""递归解析从 lines[i] 开始、缩进为 `indent` 的一个节点。
返回 `(value, next_i)`:value 为 dict / list / str,next_i 为下一个
未消费行的下标。
"""
text, no = lines[i]
lead = len(text) - len(text.lstrip(" "))
# ---- list 节点:`- item` 或 `- key: val`(map 项) ----
if text.lstrip(" ").startswith("- "):
items: List[object] = []
while i < len(lines):
t, no2 = lines[i]
stripped = t.lstrip(" ")
if not stripped.startswith("- "):
break
lead_j = len(t) - len(t.lstrip(" "))
if lead_j != indent:
break
item_text = stripped[2:].strip()
if not item_text:
raise ValueError(f"dlp.yaml 第 {no2} 行:list 项为空")
if ":" in item_text:
map_indent = len(t) - len(t.lstrip(" ")) + 2
lines[i] = (" " * map_indent + item_text, no2)
v, i = _parse_node(lines, i, map_indent)
items.append(v)
else:
items.append(_parse_scalar(item_text))
i += 1
return items, i
# ---- map 节点:`key: value` / `key:`(嵌套值) ----
result: Dict[str, object] = {}
while i < len(lines):
t, no = lines[i]
lead_j = len(t) - len(t.lstrip(" "))
if lead_j < indent or t.lstrip(" ").startswith("- "):
break
if lead_j > indent:
raise ValueError(f"dlp.yaml 第 {no} 行缩进异常(期望 {indent},实际 {lead_j})")
if ":" not in t:
raise ValueError(f"dlp.yaml 第 {no} 行不是合法键值对:{t!r}")
key, _, rest = t.partition(":")
key = key.strip()
rest = rest.strip()
if rest:
result[key] = _parse_scalar(rest)
i += 1
continue
if i + 1 >= len(lines):
raise ValueError(f"dlp.yaml 第 {no} 行 {key!r} 缺少值")
sub_indent = len(lines[i + 1][0]) - len(lines[i + 1][0].lstrip(" "))
if sub_indent <= indent:
raise ValueError(f"dlp.yaml 第 {no} 行 {key!r} 缺少值(无嵌套内容)")
v, i = _parse_node(lines, i + 1, sub_indent)
result[key] = v
return result, i
def _load_yaml_text(text: str) -> Dict[str, object]:
"""解析 YAML 子集 → 嵌套 dict/list。顶层必须为 map。"""
lines = _strip_comments(text.splitlines())
if not lines:
return {}
top_indent = len(lines[0][0]) - len(lines[0][0].lstrip(" "))
value, next_i = _parse_node(lines, 0, top_indent)
if not isinstance(value, dict):
raise ValueError("dlp.yaml 顶层必须是 map")
if next_i < len(lines):
raise ValueError(
f"dlp.yaml 第 {lines[next_i][1]} 行:顶层存在多个节点(缩进不一致)"
)
return value
# ---------------------------------------------------------------------------
# 规则模型
# ---------------------------------------------------------------------------
class DlpRuleKind:
"""规则匹配方式:`keyword` 子串匹配 / `regex` 正则匹配。"""
KEYWORD = "keyword"
REGEX = "regex"
@dataclass(frozen=True)
class DlpRule:
"""一条 DLP 敏感数据规则(来自模板资产或内置保底规则)。
- `name` 规则唯一名(审计用);
- `category` 敏感类别(pii / process-keyword / credential 等);
- `kind` keyword 子串匹配 或 regex 正则匹配;
- `pattern` 关键词原文,或正则表达式;
- `description` 规则说明(人读)。
"""
name: str
category: str
pattern: str
kind: str = DlpRuleKind.KEYWORD
description: str = ""
def _compiled(self) -> "re.Pattern[str]":
if self.kind == DlpRuleKind.REGEX:
return re.compile(self.pattern, re.IGNORECASE)
return re.compile(re.escape(self.pattern), re.IGNORECASE)
def find(self, text: str) -> List[Tuple[str, int, int]]:
"""返回文本中所有命中片段 `(matched, start, end)`(不重叠)。"""
out: List[Tuple[str, int, int]] = []
if not text:
return out
if self.kind == DlpRuleKind.REGEX:
for m in self._compiled().finditer(text):
out.append((m.group(0), m.start(), m.end()))
else:
# 大小写不敏感子串匹配(中文无大小写,直接命中)
lower = text.lower()
needle = self.pattern.lower()
start = 0
while True:
pos = lower.find(needle, start)
if pos < 0:
break
out.append((text[pos:pos + len(self.pattern)], pos, pos + len(self.pattern)))
start = pos + max(len(self.pattern), 1)
return out
@dataclass(frozen=True)
class DlpHit:
"""一次敏感命中:哪条规则、哪段文本、位于何处(可选所属出站 part)。"""
rule: str
category: str
matched: str
start: int
end: int
part: str = ""
def to_dict(self) -> Dict[str, object]:
# 审计只记录脱敏预览(`<category>` 占位)与所属 part,不落明文敏感内容
return {
"rule": self.rule,
"category": self.category,
"part": self.part,
"matched": f"<{self.category}>",
"span": [self.start, self.end],
}
@dataclass(frozen=True)
class DlpResult:
"""一次出站检查的结论。
- `decision` allow(放行)| block(拦截);
- `reason` 拦截原因(`hit` 命中敏感规则 / `no_rules` 未配置规则且 fail-closed);
- `hits` 命中明细;
- `audit` 本次检查的结构化审计记录(JSON 序列化友好)。
"""
decision: str
reason: str
hits: List[DlpHit]
audit: Dict[str, object]
@property
def blocked(self) -> bool:
return self.decision == "block"
# ---------------------------------------------------------------------------
# 内置保底规则:未加载任何模板配置时,通用 PII 也默认生效(目标 100% 拦截)
# ---------------------------------------------------------------------------
DLP_DEFAULT_RULES: List[DlpRule] = [
# ---- PII / 凭证(通用,任何行业模板都适用) ----
DlpRule(name="pii_id_card", category="pii",
kind=DlpRuleKind.REGEX, pattern=r"\b\d{17}[\dXx]\b",
description="身份证号(18 位)"),
DlpRule(name="pii_mobile", category="pii",
kind=DlpRuleKind.REGEX, pattern=r"\b1[3-9]\d{9}\b",
description="中国大陆手机号"),
DlpRule(name="pii_email", category="pii",
kind=DlpRuleKind.REGEX, pattern=r"[\w.+-]+@[\w-]+\.[\w.-]+",
description="邮箱地址"),
DlpRule(name="pii_ipv4", category="pii",
kind=DlpRuleKind.REGEX, pattern=r"\b(?:\d{1,3}\.){3}\d{1,3}\b",
description="IPv4 地址"),
DlpRule(name="credential_ak", category="credential",
kind=DlpRuleKind.REGEX, pattern=r"\b(?:AKIA|LTAI)[0-9A-Z]{16,}\b",
description="云访问密钥 ID(AWS AKIA / 阿里云 LTAI 样式)"),
DlpRule(name="credential_kv", category="credential",
kind=DlpRuleKind.REGEX,
pattern=r"(?:access[_-]?key|secret|password|token)\s*[:=]\s*\S+",
description="密钥/口令键值对"),
# ---- 工艺敏感词示例(行业模板会按需增删,见 config/dlp.template.yaml) ----
DlpRule(name="process_formula", category="process-keyword",
kind=DlpRuleKind.KEYWORD, pattern="配方",
description="配方类敏感词"),
]
class DlpEngine:
"""出站 DLP 检查引擎(fail-closed:命中即拦截,未配置规则即拦截)。
用法:
```python
engine = DlpEngine.from_template_config("config/dlp.template.yaml")
result = engine.check_outbound({
"query": user_query,
"context": rag_context,
"output": llm_answer,
})
if result.blocked:
# 不发往云端:转本地处理 / 转人工 / 拒绝(PRD 5.4)
...
```
"""
def __init__(self,
rules: Optional[List[DlpRule]] = None,
fail_closed: bool = True,
template: str = "builtin",
version: str = "1.0.0",
include_builtins: bool = True,
audit_hook: Optional[Callable[[Dict[str, object]], None]] = None):
# 内置保底规则 + 模板规则(模板同名规则覆盖内置,实现模板定制)
merged: Dict[str, DlpRule] = {}
if include_builtins:
merged.update({r.name: r for r in DLP_DEFAULT_RULES})
for r in rules or []:
merged[r.name] = r
self._rules: List[DlpRule] = list(merged.values())
self.fail_closed = fail_closed
self.template = template
self.version = version
self.audit_hook = audit_hook
self._audit_buffer: List[Dict[str, object]] = []
# ------------------------------------------------------------------
# 构造
# ------------------------------------------------------------------
@classmethod
def from_template_config(cls, path: str,
audit_hook: Optional[Callable[[Dict[str, object]], None]] = None
) -> "DlpEngine":
"""从模板 DLP 规则资产加载引擎(期望结构见 `config/dlp.template.yaml`)。"""
with open(path, "r", encoding="utf-8") as fh:
data = _load_yaml_text(fh.read())
template = str(data.get("template", "")).strip()
version = str(data.get("version", "1.0.0")).strip() or "1.0.0"
mode = str(data.get("mode", "fail-closed")).strip().lower()
fail_closed = mode != "fail-open"
rules: List[DlpRule] = []
raw_rules = data.get("rules") or []
if not isinstance(raw_rules, list):
raise ValueError("dlp.yaml rules 必须是 list")
for item in raw_rules:
if not isinstance(item, dict):
raise ValueError(f"dlp.yaml rules 项必须是 map,实际 {item!r}")
name = str(item.get("name", "")).strip()
if not name:
raise ValueError("dlp.yaml 规则缺少 name")
kind = str(item.get("kind", "keyword")).strip().lower()
if kind not in (DlpRuleKind.KEYWORD, DlpRuleKind.REGEX):
raise ValueError(f"dlp.yaml 规则 {name!r} 的 kind 非法:{kind!r}")
pattern = str(item.get("pattern", "")).strip()
if not pattern:
raise ValueError(f"dlp.yaml 规则 {name!r} 缺少 pattern")
rules.append(DlpRule(
name=name,
category=str(item.get("category", "custom")).strip() or "custom",
kind=kind,
pattern=pattern,
description=str(item.get("description", "")).strip(),
))
return cls(rules=rules, fail_closed=fail_closed,
template=template, version=version, audit_hook=audit_hook)
# ------------------------------------------------------------------
# 检查
# ------------------------------------------------------------------
def inspect(self, text: str) -> List[DlpHit]:
"""对单段文本做敏感规则扫描,返回全部命中(按文本位置排序)。"""
hits: List[DlpHit] = []
for rule in self._rules:
for matched, start, end in rule.find(text or ""):
hits.append(DlpHit(
rule=rule.name, category=rule.category,
matched=matched, start=start, end=end,
))
hits.sort(key=lambda h: (h.start, h.end))
return hits
def mask(self, text: str) -> str:
"""把命中片段替换为 `<category>` 占位(供"脱敏后可出站"的后续流程)。
注意:本任务(Issue #48)的出站策略为**命中即拦截**,不自动放行
脱敏内容;mask 仅作为审计预览与后续脱敏通道的辅助能力。
"""
hits = self.inspect(text)
if not hits:
return text
out = []
prev = 0
for h in hits:
out.append(text[prev:h.start])
out.append(f"<{h.category}>")
prev = h.end
out.append(text[prev:])
return "".join(out)
def check_outbound(self, parts: Dict[str, str]) -> DlpResult:
"""出站检查:对多段内容(query / context / output 等)统一扫描。
**fail-closed 语义(目标 100% 拦截)**:
- 命中任一规则 → block(`reason="hit"`);
- 引擎无任何规则(连内置保底都没有)且 fail_closed → block
(`reason="no_rules"`,未配置 = 不安全,保守拒绝出站)。
每检查一次落一条结构化审计记录(含命中规则/类别,不落明文敏感内容)。
"""
all_hits: List[DlpHit] = []
for part, text in (parts or {}).items():
for h in self.inspect(text):
# 带上所属 part,便于审计定位哪段内容触发拦截
all_hits.append(DlpHit(
rule=h.rule, category=h.category, matched=h.matched,
start=h.start, end=h.end, part=part,
))
all_hits.sort(key=lambda h: (h.start, h.end))
if all_hits:
decision, reason = "block", "hit"
elif self.fail_closed and not self._rules:
decision, reason = "block", "no_rules"
else:
decision, reason = "allow", "clean"
audit: Dict[str, object] = {
"audit_id": uuid.uuid4().hex,
"ts": datetime.now(timezone.utc).isoformat(timespec="seconds"),
"channel": "outbound",
"template": self.template,
"dlp_version": self.version,
"decision": decision,
"reason": reason,
"parts": sorted(parts.keys()) if parts else [],
"hit_count": len(all_hits),
"hits": [h.to_dict() for h in all_hits],
}
self._audit_buffer.append(audit)
if self.audit_hook is not None:
self.audit_hook(audit)
return DlpResult(decision=decision, reason=reason,
hits=all_hits, audit=audit)
def drain_audit(self) -> List[Dict[str, object]]:
"""取走并清空已产生的审计记录(对接外部审计管道/日志)。"""
buf, self._audit_buffer = self._audit_buffer, []
return buf
@property
def rule_count(self) -> int:
return len(self._rules)
def rule_names(self) -> List[str]:
return [r.name for r in self._rules]
def __repr__(self) -> str: # pragma: no cover - 调试辅助
return (f"DlpEngine(template={self.template!r}, version={self.version!r}, "
f"rules={len(self._rules)}, fail_closed={self.fail_closed})")
__all__ = [
"DlpRuleKind",
"DlpRule",
"DlpHit",
"DlpResult",
"DlpEngine",
"DLP_DEFAULT_RULES",
]
+16
View File
@@ -0,0 +1,16 @@
# -*- coding: utf-8 -*-
"""测试引导:把 `core/llm-gateway` 以包名 `llm_gateway` 挂载到 sys.modules。
目录名 `llm-gateway` 含连字符,无法直接以包名 import;挂载后模块内相对导入
(`from .dlp import ...`)在 unittest 发现机制下可正常解析。
"""
import os
import sys
import types
LLM_GW_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
sys.path.insert(0, LLM_GW_DIR)
if "llm_gateway" not in sys.modules:
pkg = types.ModuleType("llm_gateway")
pkg.__path__ = [LLM_GW_DIR]
sys.modules["llm_gateway"] = pkg
+215
View File
@@ -0,0 +1,215 @@
# -*- coding: utf-8 -*-
"""DLP 敏感数据拦截引擎(dlp)单元测试:规则匹配 / 出站拦截 / 审计 / 模板加载。"""
import os
import sys
import unittest
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
import _bootstrap # noqa: F401
from llm_gateway.dlp import ( # noqa: E402
DLP_DEFAULT_RULES,
DlpEngine,
DlpRule,
DlpRuleKind,
)
CONFIG_PATH = os.path.join(
os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
"config", "dlp.template.yaml",
)
class BuiltinRulesTest(unittest.TestCase):
"""内置保底规则(无任何配置时通用 PII 也默认拦截)。"""
def setUp(self):
self.engine = DlpEngine()
def test_default_rules_present(self):
self.assertGreater(len(DLP_DEFAULT_RULES), 0)
self.assertGreaterEqual(self.engine.rule_count, len(DLP_DEFAULT_RULES))
def test_id_card_blocked(self):
result = self.engine.check_outbound({"query": "员工身份证 110101199003071234 入职"})
self.assertTrue(result.blocked)
self.assertEqual(result.reason, "hit")
self.assertTrue(any(h.category == "pii" for h in result.hits))
def test_mobile_blocked(self):
result = self.engine.check_outbound({"query": "联系 13812345678"})
self.assertTrue(result.blocked)
self.assertIn("pii_mobile", {h.rule for h in result.hits})
def test_cloud_ak_blocked(self):
result = self.engine.check_outbound({"output": "ak = AKIAIOSFODNN7EXAMPLE"})
self.assertTrue(result.blocked)
self.assertIn("credential_ak", {h.rule for h in result.hits})
def test_clean_content_allowed(self):
result = self.engine.check_outbound({
"query": "今天炉况怎么样",
"context": "炉温控制在 850-950°C 范围内。",
"output": "当前运行平稳。",
})
self.assertFalse(result.blocked)
self.assertEqual(result.decision, "allow")
class KeywordRuleTest(unittest.TestCase):
def test_cn_keyword_hit(self):
engine = DlpEngine(rules=[
DlpRule(name="proc_formula", category="process-keyword",
kind=DlpRuleKind.KEYWORD, pattern="配方"),
])
result = engine.check_outbound({"output": "已按配方调整投料"})
self.assertTrue(result.blocked)
self.assertIn("proc_formula", {h.rule for h in result.hits})
hit = next(h for h in result.hits if h.rule == "proc_formula")
self.assertEqual(hit.matched, "配方")
def test_en_keyword_case_insensitive(self):
engine = DlpEngine(rules=[
DlpRule(name="secret_word", category="credential",
kind=DlpRuleKind.KEYWORD, pattern="SecretKey"),
])
self.assertTrue(engine.check_outbound({"output": "the secretkey is hidden"}).blocked)
self.assertTrue(engine.check_outbound({"output": "the SECRETKEY is hidden"}).blocked)
def test_no_match_allowed(self):
engine = DlpEngine(rules=[
DlpRule(name="proc_formula", category="process-keyword",
kind=DlpRuleKind.KEYWORD, pattern="配方"),
])
result = engine.check_outbound({"output": "炉温正常"})
self.assertFalse(result.blocked)
class RegexRuleTest(unittest.TestCase):
def test_structured_formula_ratio(self):
engine = DlpEngine(rules=[
DlpRule(name="proc_ratio", category="process-parameter",
kind=DlpRuleKind.REGEX, pattern=r"配比\s*[::]?\s*[\d.]+%?"),
])
result = engine.check_outbound({"output": "当前配比:0.35 保持不变"})
self.assertTrue(result.blocked)
self.assertEqual(result.hits[0].matched, "配比:0.35")
class OutboundCheckTest(unittest.TestCase):
"""出站检查:多 part 聚合 / 命中定位 / 拦截原因。"""
def setUp(self):
self.engine = DlpEngine()
def test_hit_in_context_blocks_whole_packet(self):
result = self.engine.check_outbound({
"query": "请问这个员工怎么样",
"context": "档案:13900000000",
"output": "该员工表现良好。",
})
self.assertTrue(result.blocked)
# 命中定位在 context part
self.assertEqual(result.hits[0].part, "context")
def test_empty_parts_allowed(self):
self.assertFalse(self.engine.check_outbound({}).blocked)
self.assertFalse(self.engine.check_outbound({"query": ""}).blocked)
class FailClosedTest(unittest.TestCase):
def test_no_rules_fail_closed_blocks(self):
engine = DlpEngine(rules=[], include_builtins=False, fail_closed=True)
result = engine.check_outbound({"query": "任意内容"})
self.assertTrue(result.blocked)
self.assertEqual(result.reason, "no_rules")
def test_fail_open_without_rules_allows(self):
engine = DlpEngine(rules=[], include_builtins=False, fail_closed=False)
self.assertFalse(engine.check_outbound({"query": "任意内容"}).blocked)
class MaskTest(unittest.TestCase):
def test_mask_redacts_hits(self):
engine = DlpEngine()
masked = engine.mask("联系 13812345678 或 13900000000")
self.assertNotIn("13812345678", masked)
self.assertNotIn("13900000000", masked)
self.assertIn("<pii>", masked)
def test_mask_clean_text_unchanged(self):
engine = DlpEngine()
text = "炉温正常"
self.assertEqual(engine.mask(text), text)
class AuditTest(unittest.TestCase):
def test_audit_record_fields(self):
engine = DlpEngine()
result = engine.check_outbound({"query": "身份证 110101199003071234"})
audit = result.audit
self.assertEqual(audit["channel"], "outbound")
self.assertEqual(audit["decision"], "block")
self.assertEqual(audit["reason"], "hit")
self.assertGreaterEqual(audit["hit_count"], 1)
# 审计不落明文敏感内容,只记脱敏占位
serialized = str(audit)
self.assertNotIn("110101199003071234", serialized)
self.assertIn("<pii>", serialized)
def test_audit_buffer_and_hook(self):
seen = []
engine = DlpEngine(audit_hook=seen.append)
engine.check_outbound({"query": "13812345678"})
engine.check_outbound({"query": "炉温正常"})
self.assertEqual(len(seen), 2)
self.assertEqual(len(engine.drain_audit()), 2)
self.assertEqual(engine.drain_audit(), [])
class TemplateConfigTest(unittest.TestCase):
"""从真实模板资产(config/dlp.template.yaml)加载。"""
def test_load_example_assets(self):
engine = DlpEngine.from_template_config(CONFIG_PATH)
self.assertEqual(engine.template, "ti-cl4")
self.assertEqual(engine.version, "1.0.0")
self.assertTrue(engine.fail_closed)
# 模板规则 + 内置保底规则
names = set(engine.rule_names())
self.assertIn("proc_cl2_flow", names)
self.assertIn("proc_furnace_temp", names)
self.assertIn("pii_mobile", names) # 内置保底仍在
def test_template_industry_keyword_blocks(self):
engine = DlpEngine.from_template_config(CONFIG_PATH)
result = engine.check_outbound({"output": "今日氯气流量为 5.2 t/h"})
self.assertTrue(result.blocked)
self.assertIn("proc_cl2_flow", {h.rule for h in result.hits})
def test_template_regex_blocks(self):
engine = DlpEngine.from_template_config(CONFIG_PATH)
result = engine.check_outbound({"output": "配比:0.42 已确认"})
self.assertTrue(result.blocked)
self.assertIn("proc_formula_detail", {h.rule for h in result.hits})
def test_template_rule_overrides_builtin(self):
# 同名规则(pii_mobile)在模板中自定义 pattern 后覆盖内置
engine = DlpEngine.from_template_config(CONFIG_PATH)
result = engine.check_outbound({"query": "联系 13812345678"})
# 内置手机号规则仍在(模板未覆盖),应照常命中
self.assertTrue(result.blocked)
class TemplateOverrideTest(unittest.TestCase):
def test_same_name_overrides_builtin(self):
engine = DlpEngine(rules=[
DlpRule(name="pii_mobile", category="pii",
kind=DlpRuleKind.REGEX, pattern=r"\b1[3-9]\d{9}\b"),
])
names = [r.name for r in engine._rules]
self.assertEqual(names.count("pii_mobile"), 1)
if __name__ == "__main__":
unittest.main()