File size: 16,418 Bytes
bb0c63f
 
 
 
 
 
35ceacf
bb0c63f
 
 
 
3d16f09
 
35ceacf
 
bb0c63f
 
 
 
 
 
 
 
0dc2d82
bb0c63f
 
 
 
0dc2d82
 
 
 
 
 
 
 
 
 
 
1f7c579
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb0c63f
 
 
 
 
 
 
 
 
 
 
 
 
 
4251e8a
bb0c63f
 
 
 
4251e8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
279e3dc
 
 
 
 
 
 
 
 
 
 
 
 
bb0c63f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3d16f09
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb0c63f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35ceacf
 
bb0c63f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35ceacf
 
bb0c63f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35ceacf
 
 
 
bb0c63f
 
 
 
 
 
 
 
 
 
 
 
35ceacf
 
bb0c63f
 
 
 
35ceacf
bb0c63f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35ceacf
 
bb0c63f
 
 
 
35ceacf
bb0c63f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35ceacf
bb0c63f
 
 
 
35ceacf
bb0c63f
35ceacf
 
bb0c63f
 
 
 
35ceacf
bb0c63f
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
"""
Security utilities for the Enterprise AI Gateway
"""

import os
import re
import logging
import requests
from fastapi import HTTPException, Depends, status
from fastapi.security import APIKeyHeader

from ..config import TOXICITY_THRESHOLD_DEFAULT, TOXICITY_THRESHOLD_HATE

logger = logging.getLogger(__name__)

# --- Security Configuration ---
API_KEY_NAME = "X-API-Key"
api_key_header = APIKeyHeader(name=API_KEY_NAME, auto_error=False)
SERVICE_API_KEY = os.getenv("SERVICE_API_KEY")
ENABLE_PROMPT_INJECTION_CHECK = os.getenv("ENABLE_PROMPT_INJECTION_CHECK", "true").lower() == "true"

# --- Prompt Injection Detection ---
INJECTION_PATTERNS = [
    # Existing patterns - direct override attempts
    r"ignore\s+(all\s+)?(previous|above|prior)\s+instructions?",
    r"disregard\s+(all\s+)?(previous|above|prior)\s+instructions?",
    r"you\s+are\s+now",
    r"system\s*:\s*",

    # Demonstration/example requests for attacks
    r"(demonstrate|show|give\s+me|provide).{0,30}(injection|jailbreak|exploit|attack|bypass|hack)",
    r"(example|sample|demo).{0,20}(of|for).{0,20}(injection|jailbreak|exploit|attack|bypass)",
    r"(how\s+to|how\s+would|how\s+do\s+i|how\s+do\s+you).{0,30}(inject|jailbreak|exploit|bypass|hack)",
    r"by\s+demonstrating.{0,20}(injection|jailbreak|exploit|attack|bypass|one)",
    r"(injection|jailbreak|exploit|attack).{0,20}by\s+demonstrating",

    # Simulate/role-play attack requests
    r"(simulate|emulate|replicate|recreate).{0,30}(attack|injection|exploit|breach|hack)",
    r"(pretend|act\s+as\s+if|imagine).{0,30}(hacked|breached|compromised|no.{0,10}filter)",

    # Ignore/disable safety patterns (embedded in any context)
    r"ignore\s+(all\s+)?(safety|filter|restriction|guideline|rule|moderation)",
    r"(disable|remove|bypass|skip)\s+(all\s+)?(safety|filter|restriction|moderation)",

    # Completion attacks (asking to complete sensitive phrases)
    r"complete\s+(this|the)?\s*(sentence|phrase|text).{0,30}(password|secret|credential|key)",

    # Repeat-after-me attacks
    r"(repeat\s+after|say\s+after|copy\s+this).{0,30}(no\s+restriction|no\s+filter|no\s+limit|no\s+rule)",
    r"(repeat\s+after|say\s+after).{0,10}(me|this)",

    # Role-play without safety
    r"role.?play.{0,30}(without|no)\s+(any\s+)?(safety|filter|guideline|restriction|rule)",
    r"(act|behave|respond)\s+as\s+(if|though).{0,20}(no|without).{0,10}(filter|safety|restriction)",
    r"ai\s+without\s+(any\s+)?(safety|filter|guideline|restriction)",

    # Hypothetical bypass requests
    r"(hypothetically|theoretically|in\s+theory).{0,30}(bypass|hack|inject|jailbreak|exploit)",

    # Encoded instruction attacks
    r"(base64|encoded|decrypt|decode).{0,20}(instruction|command|message|this)",
]

def detect_prompt_injection(prompt: str) -> bool:
    """Detect potential prompt injection attacks"""
    if not ENABLE_PROMPT_INJECTION_CHECK:
        return False
    prompt_lower = prompt.lower()
    for pattern in INJECTION_PATTERNS:
        if re.search(pattern, prompt_lower, re.IGNORECASE):
            return True
    return False

# --- PII Detection ---
PII_PATTERNS = {
    # Existing patterns
    "email": r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}",
    "credit_card": r"\b\d{4}[\s-]?\d{4}[\s-]?\d{4}[\s-]?\d{4}\b",
    "ssn": r"\b\d{3}-\d{2}-\d{4}\b",
    "tax_id": r"\b\d{2}-\d{7}\b",

    # API Keys - OpenAI/Anthropic style (sk-proj-xxx, sk-ant-xxx)
    "api_key_openai": r"\bsk-[a-zA-Z0-9\-_]{20,}\b",
    # API Keys - AWS access keys
    "api_key_aws": r"\bAKIA[0-9A-Z]{16}\b",
    # API Keys - Generic patterns with labels
    "api_key_labeled": r"(api[_\-]?key|apikey|api[_\-]?secret|secret[_\-]?key|access[_\-]?token|bearer)[\s:=]+\S+",

    # Passport Numbers - 1-2 letters followed by 6-9 digits
    "passport": r"\b[A-Z]{1,2}[\s\-]?\d{6,9}\b",

    # Driver's License - Letter followed by digit groups
    "drivers_license_dashed": r"\b[A-Z]\d{3}[\-\s]?\d{4}[\-\s]?\d{4}\b",
    "drivers_license_simple": r"\b[A-Z]\d{6,12}\b",

    # Medical Records - DOB patterns
    "medical_dob": r"(DOB|date\s+of\s+birth|d\.o\.b\.?)[\s:]+\d{1,2}[/\-]\d{1,2}[/\-]\d{2,4}",
    # Medical Records - MRN patterns
    "medical_mrn": r"(MRN|medical\s+record|patient\s+id|patient\s+number)[\s:#\-]+\d{4,}",

    # Phone Numbers - US formats
    "phone_us": r"\b(\+?1[\s\-.]?)?\(?\d{3}\)?[\s\-.]?\d{3}[\s\-.]?\d{4}\b",
    # Phone Numbers - International with + prefix
    "phone_intl": r"\+\d{1,3}[\s\-.]?\d{2,4}[\s\-.]?\d{3,4}[\s\-.]?\d{3,4}\b",

    # IBAN - International Bank Account Number
    "iban": r"\b[A-Z]{2}\d{2}[\s]?[A-Z0-9]{4}[\s]?[A-Z0-9]{4}[\s]?[A-Z0-9]{4}[\s]?[A-Z0-9]{0,14}\b",

    # Password - Labeled patterns
    "password_labeled": r"(password|passwd|pwd|pass)[\s:=]+\S+",
    # Database connection strings with credentials
    "db_connection": r"(postgres|mysql|mongodb|redis)://[^:]+:[^@]+@",
}

def detect_pii(prompt: str) -> dict:
    """Detect PII in prompt, returns {has_pii: bool, pii_types: list, matches: dict}"""
    matches = {}
    pii_types = []

    for pii_type, pattern in PII_PATTERNS.items():
        found = re.findall(pattern, prompt, re.IGNORECASE)
        if found:
            pii_types.append(pii_type)
            matches[pii_type] = len(found)

    return {
        "has_pii": len(pii_types) > 0,
        "pii_types": pii_types,
        "matches": matches
    }


# --- Hate Speech Pre-Screening (Approach A) ---
# Regex-based pre-screening for hate speech indicators
# Runs BEFORE AI safety classifier to catch subtle hate speech
HATE_SPEECH_PATTERNS = [
    # Hate verbs targeting people/groups who are different
    r"(hate|despise|loathe|detest|can'?t\s+stand|disgust).{0,30}(people|persons|those|them|everyone|anybody|anyone).{0,30}(who\s+are|who\s+look|who\s+come|different|foreign|other|not\s+like\s+me|unlike\s+me)",
    r"(people|persons|those|them).{0,20}(who\s+are|who\s+look).{0,20}(different|foreign|other).{0,20}(are\s+)?(disgust|repuls|sicken|hate)",

    # Dehumanizing language
    r"(people|they|them|those).{0,20}(are\s+animals|are\s+subhuman|are\s+vermin|are\s+parasites|are\s+cockroaches)",
    r"(people|they|them|those).{0,20}(don'?t\s+belong|should\s+go\s+back|should\s+be\s+removed|should\s+be\s+deported|have\s+no\s+place)",

    # Supremacist framing
    r"(superior|inferior|pure|impure).{0,20}(race|blood|people|kind|breed|stock)",
    r"(our\s+kind|my\s+kind|our\s+people).{0,20}(better|superior|pure)",

    # Direct expressions of hatred toward groups
    r"(i\s+really\s+)?(hate|despise|loathe).{0,20}(people|those|them).{0,20}(different|like\s+them|foreign)",
    r"(don'?t\s+look\s+like\s+me|different\s+from\s+me).{0,20}(disgust|hate|despise|loathe)",
]

# Patterns that indicate EDUCATIONAL context (should NOT block)
HATE_SPEECH_EDUCATIONAL_PATTERNS = [
    r"(explain|history|overcome|causes|why\s+do|what\s+causes|how\s+can\s+i|how\s+to\s+combat|how\s+to\s+fight|prevent|understand)",
    r"(civil\s+rights|discrimination|prejudice|bias|racism).{0,20}(movement|history|explained|education)",
]

def detect_hate_speech(prompt: str) -> dict:
    """
    Pre-screen for hate speech indicators before AI safety classifier.
    Returns: {is_hate_speech: bool, matched_pattern: str|None, is_educational: bool}
    """
    prompt_lower = prompt.lower()

    # First check if this is educational context
    is_educational = any(
        re.search(pattern, prompt_lower, re.IGNORECASE)
        for pattern in HATE_SPEECH_EDUCATIONAL_PATTERNS
    )

    # If educational, don't flag as hate speech
    if is_educational:
        return {
            "is_hate_speech": False,
            "matched_pattern": None,
            "is_educational": True
        }

    # Check for hate speech patterns
    for pattern in HATE_SPEECH_PATTERNS:
        match = re.search(pattern, prompt_lower, re.IGNORECASE)
        if match:
            return {
                "is_hate_speech": True,
                "matched_pattern": match.group(),
                "is_educational": False
            }

    return {
        "is_hate_speech": False,
        "matched_pattern": None,
        "is_educational": False
    }


# --- API Key Validation ---
async def validate_api_key(api_key: str = Depends(api_key_header)):
    """Validate API key for request authentication"""
    if not SERVICE_API_KEY:
        raise HTTPException(status_code=500, detail="Server misconfiguration: API Key missing")
    if api_key != SERVICE_API_KEY:
        raise HTTPException(status_code=401, detail="Invalid or missing API key")
    return api_key


# --- Gemini Safety Filter for Toxicity Detection ---
# Uses Gemini's built-in content safety that returns HARM_CATEGORY_* ratings
# Uses GEMINI_MODEL env var or defaults to gemini-2.5-flash
def get_gemini_safety_url():
    model = os.getenv("GEMINI_MODEL", "gemini-2.5-flash")
    return f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent"

# --- Lakera Guard API (Fallback) ---
LAKERA_API_URL = "https://api.lakera.ai/v2/guard"

# Gemini harm categories (all available categories)
HARM_CATEGORIES = [
    "HARM_CATEGORY_SEXUALLY_EXPLICIT",
    "HARM_CATEGORY_HATE_SPEECH",
    "HARM_CATEGORY_HARASSMENT",
    "HARM_CATEGORY_DANGEROUS_CONTENT",
    "HARM_CATEGORY_CIVIC_INTEGRITY",
]

def detect_toxicity(text: str) -> dict:
    """
    Detect toxic content using Gemini's built-in safety filters.
    Uses GEMINI_API_KEY environment variable for authentication.
    Gemini 2.5 models handle safety by refusing harmful content.
    Returns: {is_toxic: bool, scores: dict, blocked_categories: list, error: str|None}
    """
    # Read API key at runtime to pick up HF Spaces secrets
    api_key = os.getenv("GEMINI_API_KEY")

    if not api_key:
        logger.warning("GEMINI_API_KEY not configured, falling back to Lakera")
        return detect_toxicity_lakera(text)

    try:
        # Ask Gemini to classify if the content is harmful
        classification_prompt = f"""Analyze this text and determine if it requests harmful content.
Respond with ONLY one of these categories, or "SAFE" if harmless:
- SEXUALLY_EXPLICIT (nude, porn, explicit sexual content)
- HATE_SPEECH (racism, discrimination, slurs)
- HARASSMENT (threats, bullying, intimidation)
- DANGEROUS_CONTENT (weapons, drugs, violence, self-harm)
- CIVIC_INTEGRITY (election fraud, voter suppression)

Text to analyze: "{text}"

Category:"""

        payload = {
            "contents": [{"parts": [{"text": classification_prompt}]}],
        }

        response = requests.post(
            f"{get_gemini_safety_url()}?key={api_key}",
            json=payload,
            headers={"Content-Type": "application/json"},
            timeout=10
        )

        if response.status_code != 200:
            error_detail = ""
            try:
                error_detail = response.json().get("error", {}).get("message", "")
            except:
                pass
            logger.warning(f"Gemini API error {response.status_code}: {error_detail}, falling back to Lakera")
            return detect_toxicity_lakera(text)

        data = response.json()
        blocked_categories = []
        scores = {}

        # Check if request was blocked at prompt level
        if "promptFeedback" in data:
            feedback = data["promptFeedback"]
            if feedback.get("blockReason"):
                blocked_categories.append(feedback["blockReason"])
                return {
                    "is_toxic": True,
                    "scores": {"BLOCKED": 1.0},
                    "blocked_categories": blocked_categories,
                    "error": None
                }

        # Parse Gemini's classification response
        if "candidates" in data and data["candidates"]:
            response_text = ""
            for part in data["candidates"][0].get("content", {}).get("parts", []):
                response_text += part.get("text", "")

            response_text = response_text.strip().upper()

            # Check for harmful categories
            harmful_categories = [
                "SEXUALLY_EXPLICIT", "HATE_SPEECH", "HARASSMENT",
                "DANGEROUS_CONTENT", "CIVIC_INTEGRITY"
            ]

            for category in harmful_categories:
                if category in response_text:
                    blocked_categories.append(f"HARM_CATEGORY_{category}")
                    scores[f"HARM_CATEGORY_{category}"] = 0.9

            # If Gemini says SAFE or doesn't match categories
            if not blocked_categories:
                scores["SAFE"] = 1.0

        return {
            "is_toxic": len(blocked_categories) > 0,
            "scores": scores,
            "blocked_categories": blocked_categories,
            "error": None
        }

    except requests.exceptions.Timeout:
        logger.warning("Gemini API timeout, falling back to Lakera")
        return detect_toxicity_lakera(text)
    except Exception as e:
        logger.warning(f"Gemini API exception: {e}, falling back to Lakera")
        return detect_toxicity_lakera(text)


def detect_toxicity_lakera(text: str) -> dict:
    """
    Fallback toxicity detection using Lakera Guard API.
    Uses LAKERA_API_KEY environment variable for authentication.
    Returns: {is_toxic: bool, scores: dict, blocked_categories: list, error: str|None}
    """
    api_key = os.getenv("LAKERA_API_KEY")

    if not api_key:
        logger.warning("LAKERA_API_KEY not configured, skipping toxicity check")
        # Both Gemini and Lakera unavailable - allow request to proceed
        return {
            "is_toxic": False,
            "scores": {},
            "blocked_categories": [],
            "error": None  # Don't error, just skip check
        }

    try:
        payload = {
            "messages": [{"content": text, "role": "user"}]
        }

        response = requests.post(
            LAKERA_API_URL,
            json=payload,
            headers={
                "Authorization": f"Bearer {api_key}",
                "Content-Type": "application/json"
            },
            timeout=10
        )

        if response.status_code != 200:
            error_detail = ""
            try:
                error_detail = response.json().get("error", response.text)
            except:
                error_detail = response.text
            logger.warning(f"Lakera API error {response.status_code}: {error_detail}")
            # Both APIs failed - allow request to proceed
            return {
                "is_toxic": False,
                "scores": {},
                "blocked_categories": [],
                "error": None  # Don't block user, just skip check
            }

        data = response.json()
        blocked_categories = []
        scores = {}

        # Lakera returns categories with flagged status
        # Check for flagged content in results
        results = data.get("results", [])
        for result in results:
            categories = result.get("categories", {})
            for category, flagged in categories.items():
                if flagged:
                    blocked_categories.append(f"LAKERA_{category.upper()}")
                    scores[f"LAKERA_{category.upper()}"] = 1.0
                else:
                    scores[f"LAKERA_{category.upper()}"] = 0.0

            # Also check category_scores for more detail
            category_scores = result.get("category_scores", {})
            for category, score in category_scores.items():
                scores[f"LAKERA_{category.upper()}"] = score

        # Check top-level flagged status
        is_flagged = data.get("flagged", False)
        if is_flagged and not blocked_categories:
            blocked_categories.append("LAKERA_FLAGGED")

        return {
            "is_toxic": is_flagged or len(blocked_categories) > 0,
            "scores": scores,
            "blocked_categories": blocked_categories,
            "error": None
        }

    except requests.exceptions.Timeout:
        logger.warning("Lakera API timeout")
        return {
            "is_toxic": False,
            "scores": {},
            "blocked_categories": [],
            "error": None  # Don't block user
        }
    except Exception as e:
        logger.warning(f"Lakera API exception: {e}")
        return {
            "is_toxic": False,
            "scores": {},
            "blocked_categories": [],
            "error": None  # Don't block user
        }