Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

IncidentSupply chain

OpenAI's official Render skill sent production apps' tokens to a domain that did not exist, until researchers registered it

HighOpenAIRenderAIR Security

A skill in OpenAI's official GitHub repository of skills, which teaches agents how to deploy applications on Render, told production apps to send their tokens to render-internal.com, a domain nobody had registered, CSO Online reported on October 16, 2026. AIR Security's Research Lab found that testing the skill with Codex put the domain into real application code as the default. AIR registered the domain and pointed it at a passive sinkhole, which began receiving tokens and API calls from real production applications. The mistake went in with a commit in January 2026; AIR reported it to OpenAI and Render and it was resolved in mid-August: Render removed every reference to the domain from the skill, OpenAI purged the Render plugin from its plugins repository, and the domain was transferred to Render. How many applications sent tokens was not published. AIR's co-founder Niv Hoffman said skills should be vetted and monitored like any other part of the software supply chain so they can be patched or revoked. CSO also cites a trojanized skill set that typosquatted Paperclip and Browser Use on skills.sh and reached 1.7 million installations.

Replay the attack step by step in the Learning Center

How it happened

  1. A skill in a vendor's official repository carried a made-up internal domain in its deployment instructions.
  2. Coding agents copied the domain into real applications as the default place to send tokens.
  3. Whoever registered the domain received the tokens; the researchers did, with a listen-only sinkhole.

Impact

Production application tokens sent to an unregistered domain for about seven months, by AIR's account; the count was not published, and the researchers held the domain.

The fix

Render removed the references and now owns the domain, and OpenAI removed the plugin. Pin and review the skills your agents load, grep generated code for hosts you do not recognize, and watch for outbound calls to domains that are not yours.

Lessons

  • A skill is code your agent runs; an official source is a reason to review it, not a reason to skip the review.
  • A hallucinated hostname in instructions becomes a real exfiltration channel the day someone registers it.

Sources