Pull to refresh
Logo
Two million bait prompts reveal theft, logging and resale across the AI supply chain

Two million bait prompts reveal theft, logging and resale across the AI supply chain

New Capabilities

A five-month canary-token test through routers, resellers and model hosts finds breaches at three providers, prompt logging by middlemen and privacy flags that did not hold

Overview

Updated 2 minutes ago

Dragos, an industrial cybersecurity firm, spent five months pushing about two million prompts through popular AI services. Each prompt carried a fake Amazon Web Services key, a private URL and a hostname that had never existed before. Anyone who touched one would leave a trace.

The traces came back ugly. Attackers copied requests at three hosting providers, and resellers read prompts straight from built-in logs. One Chinese lab's systems fetched the secret URLs from more than half the prompts sent to it, including about two-thirds marked zero-retention.

Why it matters

Prompts sent through routers and resellers can be logged, resold or stolen, and zero-retention flags may not stop it.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

2M
Prompts seeded with unique canary credentials
Sent to widely used AI services between July and November 2026.
3
Hosting providers where attackers copied customer requests
Dragos says none of the three knew about the breach before being told.
~200
AWS bait keys stolen from one provider in late October
Every request sent during the affected windows was copied, including some the provider rejected.
67%
Zero-retention prompts still fetched by one lab's scraper
Roughly two-thirds of such requests sent to the lab between July and October.
24 days
Gap between a prompt and its link appearing in a Feishu chat
The full link was posted even though the request carried a zero-retention flag.

Voices

Curated perspectives — historical figures and your fellow readers.

Sign up to generate historical perspectives on this story.

People Involved

Organizations Involved

Timeline

July 2026 November 2026

6 events Latest: Today
Tap a bar to jump to that date
  1. Ars Technica reports the findings

    Today Report

    Report details the breaches and OpenRouter's response to a related access report.

  2. Dragos publishes the study

    Report

    Findings cover three provider breaches, reseller logging and ignored privacy flags.

  3. Chinese lab scrapes zero-retention prompts

    Data handling

    Lab systems fetch secret URLs from more than half the prompts sent July through October.

  4. Largest breach drains AWS bait keys

    Breach

    A couple hundred keys sent in short windows were all copied, some from rejected requests.

  5. Canary seeding begins

    Research

    Dragos starts sending prompts carrying unique URLs, hostnames and AWS keys.

  6. First provider compromise observed

    Breach

    Earliest incident dates to mid-July; requests copied at the provider's front door.

Scenarios

1

One of the three breached providers confirms the theft

Possible Resolves by Q1 2027

Discussed by: Dragos, which says it notified each provider and that none knew of the breach beforehand

Dragos says attackers lifted real customer requests alongside the bait. If one of the three providers concludes that customer credentials or prompt text was taken, US state breach-notification laws would force a public filing. A disclosure could also arrive as a security advisory if the provider decides to get ahead of it. None of the three has named itself so far.

2

A privacy regulator opens a formal inquiry into prompt retention

Possible Resolves by Q3 2027

Discussed by: Security researchers and privacy commentators reacting to the zero-retention findings

The study's sharpest finding is that a zero-data-retention flag did not prevent scraping or storage. That claim is testable, and regulators in the United States, Britain and the European Union have authority over whether privacy representations match practice. An inquiry would likely start with requests for internal routing and logging records from a router or a provider rather than a fine.

3

Routers enforce zero-retention routing end to end

Unlikely Resolves by Q2 2027

Discussed by: Dragos, which recommends auditing routing and verifying where prompts actually travel

Routers already offer a zero-data-retention toggle, but the study found requests flagged that way reaching providers that ignored it. A fix means refusing to route flagged requests to providers without matching controls, and publishing which upstreams qualify. That is a product and contract change, not a research problem, so it depends on router operators deciding the lost supply matters less than the liability.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

2010

Google Street View Wi-Fi collection (2010)

Google's Street View cars logged unencrypted Wi-Fi payload data, including emails, passwords and full web addresses, in more than 30 countries while photographing streets. The company said a single engineer's experimental code had been included by mistake.

Then

Privacy regulators in several countries opened investigations. The US Federal Communications Commission fined Google $25,000 in 2012 for delaying its inquiry.

Now

Google stopped collecting Wi-Fi payload data from its cars and added privacy review steps to product launches.

Why this matters now

Data gathered incidentally, in transit, by a system nobody thought of as a collection tool. Prompt logs at an inference provider have the same shape.

August 2006

AOL publishes search logs (2006)

AOL released a file of roughly 20 million search queries from about 650,000 users for research purposes. Queries were keyed to numeric IDs instead of names.

Then

Reporters identified individuals within days from the detail in their searches. AOL pulled the file and its chief technology officer resigned.

Now

The episode became the standard caution about so-called anonymized data and fed pressure on search engines to shorten how long they kept logs.

Why this matters now

Prompts carry as much identifying detail as search queries, and any provider that logs them holds the same kind of record.

July 2018

Gmail third-party app access (2018)

The Wall Street Journal reported that outside app developers could read the Gmail messages of users who had granted permission through a consent screen. Google said it vets developers and that users approve access.

Then

Google defended its review process and faced congressional questions about third-party access to email.

Now

Google restricted which apps could request Gmail data, narrowing the set of developers able to read message content.

Why this matters now

The intermediate layer that customers never see is where content gets read. Inference resellers occupy exactly that position.

Sources

(5)