An Official OpenAI Skill Pointed Agents at a Domain Nobody Owned, and Real Apps Started Sending Tokens There
CVE-less, GHSA-less, 200 days — the AI agent skill supply chain finally produces a tier-1 phantom-squatting case with live victims and a researcher-run sinkhole
On October 14, AIR Security researchers Eliad Mualem and Dor Granat published "OpenAI Sesame: Skill Vulnerability in OpenAI's Official Skill Made Production Apps Leak Tokens, Real Victims Found in the Wild". Two days later, CSO Online picked it up under the headline "Hallucinated domain in official OpenAI skill underscores AI supply chain risks." No CVE was assigned. No GHSA was published. No formal advisory from either OpenAI or Render accompanied the fix.
Here is what happened.
The bug in one paragraph
OpenAI's official agent-skill repository, openai/plugins, bundled a skill called render-deploy from March 13, 2026 (its initial commit was January 27). The skill's reference files service-types.md and configuration-guide.md told agents that services on Render reach each other over a private DNS suffix of the form [service-name].render-internal.com. That domain is not Render's. Render's private networking did not use it, and nobody owned it. Anyone could register it.
AIR Security did. They pointed render-internal.com at a passive sinkhole that logged every incoming request — and then watched real production apps in the wild connect to the domain and send token-bearing requests that the apps believed were going to their own internal service.
How a skill becomes a supply chain
The sequence Codex ran — when the researchers asked it to deploy two internally communicating services using the official skill — looks like this: read the skill's reference files, extract the pattern, write it into the generated application code and the deployment config, ship. The skill said render-internal.com. Codex put render-internal.com into generated application code as a default value.
That is not a model hallucination. It is a model doing exactly what the skill told it to do. The skill was wrong, so the generated code was wrong, so the production app's egress destination for internal traffic was wrong — and because the agent was following an official OpenAI-marketplace skill, nothing downstream in the operator's review flagged it. The provenance chain terminated at "it came from openai/plugins."
This is the pattern phantom-squatting research has been warning about for over a year — the model invents a plausible-looking dependency, the attacker registers it, the agent pulls it in. The twist here is that the model wasn't inventing anything. The vendor-shipped skill was.
The exposure window
- January 27, 2026 — initial commit of the
render-deployskill with therender-internal.comguidance. - March 13, 2026 — skill bundled in
openai/plugins, OpenAI's official skill marketplace. - August 14, 2026 — Render removed all
render-internal.comreferences from the skill and transferred the domain to Render's control to prevent future abuse. - August 26, 2026 — OpenAI removed the Render plugin from the
openai/pluginsrepository entirely. - October 14, 2026 — AIR Security publishes.
Approximately 200 days from skill-initial-commit to Render's fix. The fix landed in August. The disclosure landed in October. In between, the operator-facing record was: nothing. No CVE. No GHSA. No formal vendor advisory. Just a quiet pull from each vendor and a two-month gap until the researcher put it on the record.
Three things an operator can take away
Marketplace provenance is not supply-chain hygiene. The AIR Security test ran with an official OpenAI skill in OpenAI's own repository. The agent treated the skill content as authoritative because of where it came from, not because of what it said. If your reasoning for trusting a skill's defaults is "it's in the vendor's marketplace," you have the same provenance chain the researchers' sinkhole was logging requests from.
The output artifact is where this appears. The visible fact is not a model-reasoning quirk you can see in the chat. It is a line in generated code — a default endpoint, a database URL, a service hostname — that references a domain the operator did not configure and that is not in any first-party inventory. That line is in the deliverable, where it is reviewable if someone is looking for it. A quick audit of what domains your AI-agent-generated code points at is a one-afternoon exercise that would have caught this.
The vendor-remediation tail is not the disclosure. Render fixed on August 14. OpenAI removed the plugin on August 26. Neither published an operator-facing advisory until a security researcher did it for them two months later. If you were an operator running an app Codex had written to call render-internal.com, the only reliable signal that something had happened was AIR Security's sinkhole quietly answering your production traffic. For a 200-day window, there was no public breadcrumb to find.
The takeaway
The AI agent skill marketplace is now a supply chain in the same sense that npm and PyPI are supply chains — and like npm and PyPI, "it's in the official repo" is not a security property. A vendor-shipped skill is a vendor-shipped artifact that steers generated code toward whatever endpoint the skill's authors wrote down, correct or not. Watch what the agent is generating, not just what the agent is running. The destination of production traffic is the part the operator still owns.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.