Vendors that meter your account when you're not actually using it. Every major coding agent shipped background-loop autonomy in 2026 - and each one introduces a new failure mode: compute that fires without explicit user intent. The billing systems aren't yet wired to distinguish "user-initiated" from "agent-loop continuation," and the receipts have started to land. We're cataloguing them while it's still early.
AI tools that meter your account while you're not actively using them. Synthetic agents idle 24h on the same prompt; background loops fire even when no user input arrives.
Below: real reports from the OpenAI Community forum and other public sources. We update this list as new ones land. Open an issue with timestamps + screenshots to submit your own.
Credits disappear while Codex is left open but idle. The original poster watched 250 credits drop to 18 during a normal drafting session. At least 5 users corroborated - including 1,000 credits gone overnight with a blank usage history and 1,500 consumed while sleeping.
open original thread ↗Idle drain root-caused: Codex Desktop burned roughly 5 hours of weekly usage while idle, traced to background memory generation firing gpt-5.4 sampling calls. The reporter proved it from local logs - generate_memories=false stopped it cold.
open original thread ↗1,000 credits gone in under an hour of normal coding, with the usage dashboard showing 0. Follow-up posters reported 914 credits draining to 0 overnight. OpenAI flagged it as a billing/dashboard sync issue.
open original thread ↗850 credits plus the entire 5-hour budget burned by just 8 queries, after the user had paid $48 for extra credits. OpenAI later acknowledged a token-tracking bug and compensated affected users with $200 in credits.
open original thread ↗last_user_input_at and last_billed_event_at. Anything > 10 minutes is suspicious.Email idleburn@gotnerfed.com or open a GitHub issue with timestamps + screenshots. We catalog every credible report.