Signum
Feed
Useful signal22 Nov 2026medium confidence

Dragos research finds at least three compromised AI inference providers stealing credentials; OpenRouter removes three providers and revokes ZDR approval for 125 endpoints

Edited by · written with AI from the linked sources (how)

Dragos published findings from about 2 million test requests with fake credentials (July-November 2026). It identified at least three compromised providers where attackers harvested credentials (hundreds of AWS keys in the largest case), resellers reading customer requests, and Zero Data Retention settings often ignored. In response, OpenRouter removed three providers, revoked ZDR approval for 125 endpoints and emailed affected users about possible third-party reads of MiniMax requests. Novita, Phala and Together have published preliminary security reports.

InfrastructureGovernanceEconomicsAccess

Entities: Dragos, HD Moore, OpenRouter, MiniMax, Novita, Phala

76Useful signal
1 source
0 primary
Was this useful?
01

What happened

Security firm Dragos sent about 2 million test requests carrying fake credentials to AI inference providers between July and November 2026. It reports at least three compromised providers where attackers harvested the fake credentials (hundreds of AWS keys in the largest case), resellers reading customer requests, and Zero Data Retention (ZDR) settings often ignored. OpenRouter has since removed three providers, revoked ZDR approval for 125 endpoints and emailed users about possible third-party reads of MiniMax requests. Novita, Phala and Together have published preliminary security reports. The coverage is The Decoder's write-up, and no link to the Dragos report itself is included, so exact methods and provider names are not verified here.

02

Why it matters

Anyone sending prompts, API keys or customer data through low-cost routers and resellers cannot assume a privacy promise such as ZDR holds along the whole chain. Developers and enterprises using routed or marketplace inference have reason to review which providers their traffic can reach, check privacy settings and add canary tokens (fake credentials that reveal misuse). OpenRouter's action is a concrete sign that at least one major aggregator accepted the findings. The scale of real-world harm is still unclear: the research was a probe with decoy credentials, and we do not know how many real users were affected.

03

What is noise

The mention of Chinese actors adds a geopolitical angle without clear public evidence in this summary of who the attackers were. "At least three" compromised providers is a floor from one firm's test, not a measured rate across the market, and the headline's alarm outpaces what is known about actual customer losses. Note too that the same intermediaries that carry the risk are named as the places where fixes are being made, so the preliminary reports from Novita, Phala and Together are self-reported.

04

Watch next

  1. 01Publication of the full Dragos report or data, naming the compromised providers and showing how attackers got in and how the ZDR failures were measured, so others can reproduce it.
  2. 02Final (not preliminary) security reports from Novita, Phala and Together, and whether OpenRouter restores ZDR approval for any of the 125 endpoints and says how many users the MiniMax notice covered.
  3. 03Confirmed cases of real customer credentials being abused (for example AWS keys used by attackers), or changes to routing and data-retention terms at other aggregators, which would show this is a market-wide problem rather than a few bad providers.

Coverage

1 story

Newsletter

Get the no-hype AI news straight to your inbox

Weekly. No spam, unsubscribe any time. Privacy

More infrastructure signals

Full feed →