If budgets dont actually stop azure openai spend, what does?
retained text
Stolen keys are the obvious way this goes wrong. microsofts DCU went after storm-2139 for using customer credentials scraped from public sources to get into azure openai, and sysdigs LLMjacking research estimated over $46K a day for a victim on claude 2 and over $100K a day with claude 3 opus. an agent stuck in a loop is the other half and theres no attacker needed for that one what most of us lean on only alerts tho. budgets dont stop consumption, cost data takes 8 to 24 hours to land and budgets get evaluated once every 24 hours. anomaly detection runs 36 hours after the day ends and the defender for AI alerts are still just alerts. pay as you go doesnt even have a spending limit you can turn on for the stolen key side its disableLocalAuth plus managed identity, enforced with the azure policy so a scraped key stops working. microsofts own doc says the gateway can keep accepting old keys for a few minutes up to several hours after you flip it so no instant cutoff. The slower part is finding everything that still calls with a key before you flip it, since anything you miss just starts failing with 401s. that list can come from the request logs on the resource, an nhi inventory like hush security or others or just a grep thru app settings and key vault refs loops are harder since nothing is stolen, so the cap has to sit in the request path and not read yesterdays bill. what azure gives you there: deployment TPM, which caps the burn rate (429) but not the total, so worst case is roughly TPM x 1,440 tokens a day APIMs llm-token-limit with a token quota per key per hour, day or month (403), but only for traffic that actually goes thru APIM project level TPM and token quotas in the foundry AI gateway, which is the same APIM underneath
document fetched 2026-11-03 00:38 · canonical https://reddit.com/r/AZURE/comments/1xlvc0z/if_budgets_dont_actually_stop_azure_openai_spend