Request blocked by provider content policy

The difference from the safety trigger: there the model answered with a refusal, here the provider did not run the request at all — it returned a moderation error (content_filter, responsible_ai_policy_violation, content_policy_violation). To you it looks like a 400/403 in production, and the user then sees a blank screen or a generic error.

Such blocks are almost always about a specific prompt or a specific user input. A push with a truncated and redacted sample gives you exactly what you need to fix it without digging through logs.

Catch the moderation error in code

The key point is to distinguish a content-policy error from other 400s. Each provider has its own code, so we collect them into one set:

import os, re, requests

NOTIFLY_URL   = os.environ["NOTIFLY_URL"]
NOTIFLY_TOKEN = os.environ["NOTIFLY_TOKEN"]

POLICY_CODES = {
    "content_filter",                    # OpenAI / Azure
    "content_policy_violation",          # OpenAI images
    "responsible_ai_policy_violation",   # Azure
    "safety",                            # Vertex / Gemini
}

def notify(title, msg, prio):
    requests.post(f"{NOTIFLY_URL}/message",
                  params={"token": NOTIFLY_TOKEN},
                  json={"title": title, "message": msg, "priority": prio},
                  timeout=5)

# rough redaction: strip e-mails, phones, long digit sequences
def redact(text: str, limit=400) -> str:
    text = re.sub(r"[\w.+-]+@[\w-]+\.[\w.-]+", "[email]", text)
    text = re.sub(r"\+?\d[\d\s()-]{7,}\d", "[phone]", text)
    text = re.sub(r"\b\d{6,}\b", "[num]", text)
    return text[:limit]

Now wrap the provider call:

import openai

def ask(user_input: str):
    try:
        return client.chat.completions.create(
            model="gpt-4o",
            messages=[{"role": "user", "content": user_input}],
        )
    except openai.BadRequestError as e:
        body = getattr(e, "body", {}) or {}
        code = (body.get("code")
                or body.get("error", {}).get("code")
                or "")
        if code in POLICY_CODES:
            notify("🚫 Content-policy: запрос заблокирован",
                   f"Provider: openai\nCode: {code}\n\n"
                   f"Ввод (обезличен):\n{redact(user_input)}",
                   prio=7)
        raise

priority=7 — this is not a security incident but a “fix the prompt/input” signal, so a normal loudness rather than a “ten”.

Deduplication: the same prompt fails hundreds of times

If it is your system prompt that is broken, not a one-off user input, the error will fire on every request. You want to send the first case and then a periodic aggregate — the common pattern:

import time
state = {"first_ts": 0, "count": 0}
WINDOW = 3600

def maybe_alert(code, sample):
    now = time.time()
    if now - state["first_ts"] > WINDOW:
        state.update(first_ts=now, count=0)
        notify("🚫 Content-policy (первый за час)",
               f"Code: {code}\n{sample}", prio=7)
    state["count"] += 1
    if state["count"] in (25, 250):
        notify(f"🚫 Content-policy ×{state['count']}",
               "Похоже, задет ваш системный промпт или шаблон — а не разовый ввод.",
               prio=9)

Other providers

• Anthropic — stop_reason == "refusal" in the response (the request ran but the model refused), or a 400 with error.type == "invalid_request_error".
• Vertex / Gemini — the response has promptFeedback.blockReason and safetyRatings; the block does not raise an exception, you must check it in the body.
• Providers without a webhook — if moderation arrives only by email, wrap the mailbox in an Email Inbox and alert on the emails.

What to put in the alert text

• the provider and the exact policy code;
• the truncated and redacted input (no raw PII in a push!);
• which category fired, if the provider returns it (violence / hate / …);
• the session/request id, to find the full log in S3/Sentry.