"Read by people": HD Moore caught AI providers keeping zero-retention prompts

From July to November 2026 HD Moore of Dragos sent about two million prompts with planted secrets to AI services. The secrets exposed breaches at three hosting providers and links from prompts marked for zero retention that were fetched days later.

Console of the Dragos prompt-seeding service: new incidents, seeded prompts and spending per day in November 2026, split by router
Picture: Dragos
In this material
  1. Breaches and resellers
  2. Zero retention in practice
  3. The notice from OpenRouter

Moore, who created the Metasploit penetration testing framework and joined Dragos when it bought his company runZero in September, published the results on 19 November. The prompts carried real AWS keys to empty, fully logged accounts, unique links and passwords. Any use of one meant someone along the chain had seen the prompt.

Zero Data Retention is supposed to mean something. In practice, these prompts were retained, shared with other companies, and read by people.

HD MooreChief research officer, Dragos; creator of MetasploitDragos blog, November 19, 2026

Breaches and resellers

Attackers were stealing data from three hosting providers, in one case since mid-July, and none of the three knew until Dragos told them. At the largest, every request sent in a few windows of late October was stolen, even ones rejected before a model ran. At the other two, some stolen requests carried the zero retention flag. The keys were tested, some within hours, and used for tens of thousands of AWS calls.

Resellers on an inference marketplace needed no break-in: they listed a model at the lowest price, switched on request logging and waited. A day and a half after two prompts went through one seller, someone on a home connection opened both secret links and started guessing user names.

Zero retention in practice

The Chinese lab’s scraper usually came two or three days after a request, also for prompts sent to US hosts that resell its models. One large US provider crawled a secret link more than two months after a single search request. A link from another prompt with the zero retention flag was posted in a Feishu work chat 24 days later and opened from three desktops and three phones in Beijing and Shanghai within two minutes.

The notice from OpenRouter

An email OpenRouter sent customers on 21 November, posted on Hacker News and reported by Ars Technica, says a security researcher placed unique test URLs in prompts sent to MiniMax models from July to October. A third party opened them 40 to 130 hours later, whether MiniMax served a request or AtlasCloud, Novita or StreamLake forwarded it. OpenRouter removed the Zero Data Retention label from those endpoints while it investigates. By the email’s account, MiniMax told StreamLake that content moderation was on before 22 October and zero retention since. MiniMax has not given OpenRouter its own account or commented in public. The email does not name the researcher.

Moore advises hosting models yourself where you can afford it, otherwise using one first-tier provider directly, and seeding your own canaries. The big providers’ own assistants and their prices are on the hub’s tools page.

From the pulse

All news →
  1. Tools

    Grok Build share links exposed the data of thousands of users, SpaceXSI's own apps among them

    Security firm Claroc found that the links for sharing apps made with the coding agent could be enumerated; they opened business plans, medical results and employee ID numbers. 24 of the exposed apps belonged to SpaceXSI, the merged SpaceX and xAI.

    Source: arstechnica.com

  2. Law

    US requires a licence to ship AI chips to data centres that may serve China

    An interim rule of the Commerce Department covers exports, re-exports and transfers of AI chips and servers to data centres anywhere in the world if they may provide compute to companies in China or in other countries of the D:5 group, Russia among them. Until now a licence was needed only when the buyer was based in one of those countries or owned by a company from there.

    Source: techcrunch.com

  3. Tools

    ChatGPT Wallet lets agents pay for purchases after the user approves each one

    OpenAI says agents in ChatGPT Work, Codex and dots can now complete purchases with the user's approval and the protections of their existing cards.

    Source: openai.com In the catalog: ChatGPT

More in this section

All materials →
News

"Working with Claude": Levent Alpöge's Riemann hypothesis proof passes Lean checks

On 20 November 2026 Anthropic mathematician Levent Alpöge released a proof of the Riemann hypothesis written in the Lean proof language. A day later the Lean and Mathlib communities certified the statement and the proof; a version written for mathematicians to read has not been published.

News

Shinji Mikami wishes AI would "all just go away"; Kim Hyung-tae says games can't avoid it

At G-CON 2026 in Busan on 19 November, the director of the first Resident Evil and the head of Shift Up described how their studios use AI. Mikami keeps it away from design; Kim said game development without it no longer exists.

News

Cliff Bleszinski on Big Walk's AI backlash: "People really need to get over themselves"

House House said an AI chatbot helped with about 10 lines of Big Walk's code after IGF jurors refused to vote for games made with generative AI. Negative Steam reviews followed, and the Gears of War designer took the studio's side.