Okta Tracks AI Token Resellers Using Cloud Credits
Okta Threat Intelligence reported on August 4 that gray-market AI proxy services are using fraudulent sign-ups and promotional cloud credits to sell discounted access to frontier models. One service advertised prices at 5% to 15% of official rates and exposed a status endpoint showing 881 users; a separate case study logged more than 105,000 suspected automated sign-up attempts. Buyers risk abrupt service loss and exposure of prompts to intermediaries.
Okta Threat Intelligence published an investigation on August 4 into gray-market services that resell access to frontier AI models through proxy APIs. The researchers linked some of those services to fraudulent registrations, free trials and promotional cloud credits, while carefully distinguishing observed infrastructure from claims made by the operators themselves.
What Okta found
Okta said it found advertisements for more than half a dozen services on underground forums and messaging platforms. One, called Poison Claude, advertised Anthropic-compatible API access at 5% to 15% of official per-token prices. The service told customers to redirect Claude Code to its proxy endpoint and claimed that requests were routed through a pool of accounts funded with AWS bonus credits. Okta described fraudulent registration as the probable way those accounts obtained Bedrock access, rather than presenting that mechanism as independently proven for every account.
A configuration mistake exposed a Poison Claude status route showing 881 total users and 872 active users when Okta checked it. The researchers also traced an unprotected related endpoint to a Hostinger server in Mumbai, even though the main site sat behind Cloudflare.
A second service, Ecomagent, sold discounted access to Claude models and GPT Codex. Its documentation showed a response identifier consistent with Google Vertex infrastructure, and the operator claimed to rely on startup credits. Okta noted an important limitation: a later test did not reproduce that Vertex-style identifier. The evidence therefore supports a reported proxy operation and an operator claim about credits, not a blanket conclusion that every response came through Google infrastructure.
Okta separately analyzed abuse against an unnamed AI video service. From June 1, 2023, through July 1, 2026, its logs recorded more than 105,000 suspected automated sign-up attempts from 251 IP addresses associated with bot activity. Disposable email domains, VPNs and proxy networks supported the assessment that synthetic identities were being created at scale.
The risk for teams using cheap proxies
A proxy operator can see the prompts and responses passing through its infrastructure. That creates obvious exposure for source code, credentials, research data and other sensitive material. Customers also face model substitution, sudden account shutdowns and loss of service when cloud providers detect abuse.
Okta said it notified Cloudflare, Anthropic, AWS and Google Cloud before publication. The practical lesson is narrow but important: discounted access through an unofficial AI gateway transfers control of both identity and data to an intermediary whose upstream accounts may disappear without notice.
Key Points
- 1Okta documented gray-market AI proxies that claimed to use promotional cloud credits and fraudulent registrations to undercut official model prices.
- 2Poison Claude exposed a status endpoint showing 881 users, while a separate AI-video case study recorded more than 105,000 suspected automated sign-up attempts from 251 IP addresses.
- 3Unofficial proxy customers risk prompt exposure, model substitution and abrupt service loss if upstream providers detect account abuse.
Scoring Rationale
Okta provides first-party threat-intelligence observations, concrete service and signup telemetry, and clear operational risk for AI users and cloud providers. The impact is meaningful but bounded because some credit-abuse mechanisms remain operator claims or researcher assessments, and the documented proxy services each had fewer than 1,000 visible users.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
