Dragos AI canary study: 3 AI providers breached, prompts stolen
Dragos sent nearly 2 million AI prompts with bait secrets. Three providers were breached and resellers logged prompts. What to change in your AI stack.
Your prompts travel further than your contract says. In a Dragos study published November 19, HD Moore, the company's chief research officer, describes how his team sent nearly two million AI prompts. Each prompt held a unique bait secret: a trackable URL, a password, or an AWS key. Then they watched where the secrets showed up. The result: at least three AI providers were breached, and none of them knew until Dragos told them.
What actually happened
Per the Dragos write-up:
- Scale. Close to two million prompts over five months, from July through November 2026. Every bait secret was unique, so each leak traced back to one request and one provider.
- Three breaches. The earliest started in mid-July. In the largest, attackers took "a couple hundred" of the bait AWS keys. Every request Dragos sent during the affected windows was stolen, including requests the provider rejected before any model ran. Dragos's conclusion: the attacker was copying traffic at the provider's front door.
- Fast abuse. Some keys were tested within hours. Attackers listed S3 buckets, searched Secrets Manager, and tried to turn on Bedrock models and add IAM users. The bait keys made tens of thousands of AWS API calls.
- Resellers logged prompts. On one inference marketplace, sellers only had to list a model at the lowest price, turn on request logging, and wait. No exploit needed. The same model through a different seller stayed clean. The leak followed the seller, not the marketplace.
- Privacy flags did not hold. Some stolen requests were marked zero data retention (ZDR). Dragos says providers often ignore these settings, on purpose or through bugs.
- Cross-border hops. Some prompts sent to US-based providers reached Chinese AI organizations. One seeded link turned up in an internal company chat in China.
Dragos does not name the breached providers, the marketplace, or the resellers.
Why it matters for your business
Most small teams pick an AI API by price. The cheapest route is often a router or reseller that forwards your request through hops you never see. Dragos is blunt about it: every hop in the chain can see the whole prompt and the whole response.
Now think about what goes into your prompts. Agent context with credentials. Customer emails. Pasted config files. A support bot that reads order records. If any hop is compromised, all of it is gone, and you will not get a breach notice. These providers did not know.
What we would change this week:
- Keep secrets out of prompts. Agents should call tools through a gateway that holds the keys, not paste them into context.
- Know your route. List every AI endpoint your apps call. If it is a router, find out which upstream providers it uses. A "no logging" checkbox is not proof.
- Go direct for sensitive work. One contract, one privacy policy, one company to hold to it.
- Seed your own canaries. Put a fake key in your agent context and alert on any use. That is how you find out first.
Key takeaways
- Dragos sent nearly 2 million prompts with unique bait secrets from July to November 2026
- At least three AI providers were breached; none knew until Dragos told them
- Stolen AWS bait keys were tested within hours and made tens of thousands of API calls
- Marketplace resellers logged prompts with no exploit needed
- Zero data retention flags did not stop stolen or retained prompts
Not sure where your prompts go? We map every AI endpoint your business calls, move credentials out of prompts and behind a gateway, and set up canary alerts so you hear about a leak before the bill does. See what we build or book an AI stack review.
Sources: Dragos: Three breaches, endless leaks: Your AI supply chain needs canaries.
- #ai-security
- #dragos
- #api-keys
- #ai-routers
- #prompt-privacy
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
OpenRouter drops 3 providers, pulls ZDR from 125 endpoints
OpenRouter off-boarded three inference providers and removed 125 endpoints from zero data retention after Dragos canary research. What to check in your stack.
Read itOpenAI Performance Factory: Codex agents debug, engineers merge
OpenAI's Performance Factory ran 1,700+ Codex investigations into ChatGPT slowdowns. 70% of audited root causes held up; about 15 agent PRs shipped as written.
Read it