There is a resale market for stolen LLM tokens, and an unprotected endpoint is its supply

Simon Willison surfaced on 26 July 2026 an investigation by Matt Lenhard, published on Vectoral in June 2026, mapping a four-tier and largely Chinese-language market that resells LLM API access at up to 97.8% off list — one sampled package offered about $3,333 of Anthropic credit for 425 RMB (~$58), roughly 13 cents of official usage per dollar spent, across ten leading relays drawing a combined 3.6 million monthly visits. Supply comes from bulk-registered accounts, virtual cards and unprotected endpoints, pooled through the open-source OpenAI-compatible gateways one-api and new-api; buyers want cheap inference, geographic workarounds, or tokens to distil domestic models. The practitioner action is unglamorous and immediate: hard per-key spend caps, concurrency limits, and friction on fresh accounts.

What it means Any team that exposes an LLM-backed endpoint without a spend cap is now a supplier to a priced, tooled market that actively hunts for one.