安全、合规与负责任的 LLMOps 运营

1# src/security/input_rail.py
2import re
3from typing import Tuple
4
5class PIIScrubber:
6    def __init__(self):
7        # Regex for email addresses
8        self.email_pattern = r'[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}'
9        # Regex for phone numbers (simple US format)
10        self.phone_pattern = r'\b\d{3}[-.]?\d{3}[-.]?\d{4}\b'
11
12    def scrub(self, text: str) -> Tuple[str, bool]:
13        """
14        Returns (scrubbed_text, was_modified)
15        """
16        scrubbed = text
17        was_modified = False
18        
19        # Redact Emails
20        if re.search(self.email_pattern, scrubbed):
21            scrubbed = re.sub(self.email_pattern, "<EMAIL_REDACTED>", scrubbed)
22            was_modified = True
23            
24        # Redact Phones
25        if re.search(self.phone_pattern, scrubbed):
26            scrubbed = re.sub(self.phone_pattern, "<PHONE_REDACTED>", scrubbed)
27            was_modified = True
28            
29        return scrubbed, was_modified

1# src/security/output_rail.py
2
3class OutputValidator:
4    def __init__(self):
5        self.banned_terms = ["competitor_product_x", "confidential", "internal_use_only"]
6
7    def validate(self, response_text: str) -> str:
8        """
9        Checks output for banned terms. Returns the text or a canned refusal.
10        """
11        lower_text = response_text.lower()
12        
13        for term in self.banned_terms:
14            if term in lower_text:
15                # Log this security event!
16                print(f"SECURITY ALERT: Model generated banned term '{term}'")
17                return "I apologize, but I cannot answer that question."
18        
19        return response_text