Secrets in LLM Logs

LLMs like to quote user input in their responses. If a user accidentally (or intentionally) sends their OPENAI_API_KEY — it will end up in your logs. That’s a compliance incident, plain and simple.

The simplest detector to run over any payload:

import re, os, requests

PATTERNS = [
    r"sk-[A-Za-z0-9]{20,}",                 # OpenAI / Anthropic style
    r"AKIA[0-9A-Z]{16}",                    # AWS Access Key
    r"ghp_[A-Za-z0-9]{30,}",                # GitHub PAT
    r"xox[baprs]-[A-Za-z0-9-]{10,}",        # Slack
    r"AIza[0-9A-Za-z\-_]{35}",              # Google API key
    r"eyJ[A-Za-z0-9_\-]{20,}\.[A-Za-z0-9_\-]{20,}\.[A-Za-z0-9_\-]{20,}",  # JWT
]
RX = re.compile("|".join(PATTERNS))

def scan(text: str, where: str):
    m = RX.search(text or "")
    if m:
        token = m.group(0)
        push(f"🔑 LEАК в {where}",
             f"Похожий на секрет токен: {token[:6]}…{token[-4:]}\n"
             f"Длина: {len(token)} символов\n\n"
             "Немедленно ротируйте ключ и удалите запись из логов.",
             priority=10)
        return True
    return False

def push(t, m, p):
    requests.post(f"{os.environ['NOTIFLY_URL']}/message",
                  params={"token": os.environ["NOTIFLY_TOKEN"]},
                  json={"title": t, "message": m, "priority": p}, timeout=5)

Run scan() on the user-input before sending it to the LLM (reject), and on the model output before writing it to logs (mask + alert).

Related recipes

• Safety / prompt injection triggered.