SafetyRisk, alignment & guardrails
Dragos canary study finds three AI inference providers breached and ZDR privacy flags ignored across AI routing chains
Dragos researcher HD Moore seeded nearly 2 million AI prompts with unique bait secrets. The bait exposed breaches at three inference providers, resellers logging prompts, and a Chinese lab fetching URLs from requests flagged Zero Data Retention. OpenRouter then warned MiniMax users.
From the Ars Technica · AI article of November 20, 2026 and cited sources: Attackers inside at least three inference hosting providers copied every request sent during certain late-October windows, including requests rejected before any model ran. [1] [2]
SOURCES[1] Researcher plants 2m canaries into the AI supply chain. A… arstechnica.com[2] Dragos: Three breaches, endless leaks: Your AI supply… dragos.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Attackers inside at least three inference hosting providers copied every request sent during certain late-October windows, including requests rejected before any model ran. [1] [2]
- Original source
- Ars Technica · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 20, 2026 article published
- Covered
- Nov 21 edition
- Evidence
- 8 sources across 8 sites
- 1Researcher plants 2m canaries into the AI supply chain. A…arstechnica.com
- 2Dragos: Three breaches, endless leaks: Your AI supply…dragos.com
- 3TheCyberSecGuru: OpenRouter MiniMax Prompts Readthecybersecguru.com
- 4Hacker News: Tell HN: OpenRouter prompts to Minimax may…news.ycombinator.com
- 5Cloud Security Alliance: Fourth-Party AI Subprocessor Risklabs.cloudsecurityalliance.org
- 6OpenRouter Data Processing Agreementopenrouter.ai
RisingHow it’s measured
Hacker News · 25 likes · 5 comments
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

Attackers inside at least three inference hosting providers copied every request sent during certain late-October windows, including requests rejected before any model ran. They reused about 200 stolen AWS keys in tens of thousands of API calls, probing S3, Secrets Manager and parameter stores, trying to enable Bedrock models and adding their own IAM users. Keys from two of the providers were used from the same rented server. [1] [2]
Zero Data Retention labels often did not hold. A Chinese lab fetched secret URLs from more than half of the prompts it received, including about two-thirds of requests marked ZDR or no data collection, usually 2–3 days later. OpenRouter's customer notice separately confirms URL access 40–130 hours after requests to MiniMax models and removes ZDR labels from those endpoints. MiniMax attributes the retention to content moderation. Dragos does not name the lab, so linking it to MiniMax is an inference. [1] [3] [4]
Multi-hop routing is the structural weakness. Routers such as OpenRouter and Vercel read prompts in plaintext to choose a provider, and a privacy flag binds only the first hop. One marketplace added about 1,000 tokens per prompt, which revealed an undisclosed routing step. OpenRouter's data processing agreement says its ZDR list reflects each provider's stated policy, not guaranteed compliance. [1] [2] [6]
Moore ranks the defences: self-host sensitive inference first, then contract directly with a first-tier provider, and otherwise pin upstream providers rather than accept default router pools. Practitioners should also scrub credentials from prompts, seed their own canaries in agent contexts, and rotate any secrets sent to MiniMax models between June and October 22. [1] [2] [7]
Reused canaries show secrets leaking through multi-hop AI routing despite privacy labels. Router ZDR settings therefore work as policy statements, not technical controls.
Read the full assessment
Teams handling credentials or regulated data likely need direct contracts, pinned providers and credential scrubbing.
Executive brief
Over five months, Dragos researcher HD Moore sent nearly 2 million prompts to AI services. Each prompt carried a unique, never-used secret. Attackers inside at least three inference hosting providers stole every request sent during certain windows in late October, and some of those requests had been rejected before any model ran. They then used the bait AWS keys in tens of thousands of API calls (Ars Technica). Separately, a Chinese lab fetched secret URLs from requests marked Zero Data Retention (ZDR), meaning the provider is not supposed to store them. For anyone sending confidential data through AI routers, privacy flags offered little real protection.
What changed and event timeline
Canary seeding begins
Moore starts placing AWS keys, SSH passwords, URLs, DNS names and git repository references into prompts, about 2,000 routes in all. Dragos dates the first incident to mid-July ().
MiniMax changes its handling
According to MiniMax, content moderation was switched on for its endpoints before this date, and ZDR has been enabled since. OpenRouter's customer notice covers requests from June 6 to October 22 ().
- Late Oct 2026
Three provider breaches
Attackers reuse roughly 200 stolen AWS keys, testing some within hours. They search S3, Secrets Manager and parameter stores, try to enable Bedrock models and add their own IAM users ().
AI attack tooling appears
One attacker deploys ARTEX, which Dragos describes as an LLM-driven penetration-testing framework ().
Dragos publishes
The write-up reports breaches, resellers taking prompts through logging, and ignored privacy flags. About 12 providers are still investigating ().
OpenRouter warns customers
OpenRouter emails customers that requests sent to MiniMax models "may have been read" and removes the ZDR labels from those endpoints ().
Capabilities and access
This is a security measurement study, not a model release. No specific model versions are documented.
Read the full section
This is a security measurement study, not a model release. No specific model versions are documented. The services involved:
- AI routers (gateways): OpenRouter and Vercel, which read each prompt in plaintext to choose a provider and can chain to each other (Ars Technica).
- Hosts that forwarded MiniMax-model traffic to MiniMax: AtlasCloud, Novita and StreamLake (Hacker News).
- Canary tooling: Dragos cites Thinkst Canarytokens as its initial detection tool (Dragos).
Technical analysis for researchers and developers
- Design: each secret went into exactly one prompt sent to one provider, so any reuse points back to a single route.
- Attribution signals: fetch IPs, HTTP client fingerprints, timestamps, error-envelope analysis to identify software layers, and CloudTrail logs (Dragos).
- Hidden hops: one marketplace silently added about 1,000 tokens per prompt, which revealed an undisclosed routing step (Dragos).
Read the full section
- Design: each secret went into exactly one prompt sent to one provider, so any reuse points back to a single route. Sending the same model through different sellers separated reseller logging from model-host behaviour (Ars Technica).
- Attribution signals: fetch IPs, HTTP client fingerprints, timestamps, error-envelope analysis to identify software layers, and CloudTrail logs (Dragos).
- Hidden hops: one marketplace silently added about 1,000 tokens per prompt, which revealed an undisclosed routing step (Dragos).
- Reproducibility: no dataset or code has been released. Dragos says canaries cannot always pinpoint where in a chain a leak happened.
Claims and evidence
- Three breached providers, front-door copying, one shared adversary: secrets from two of the providers were used from the same rented server (Ars Technica).
- Chinese lab scraping: URLs fetched from more than half of prompts, including about two-thirds of ZDR or "no data collection" requests, typically 2–3 days later (Ars Technica).
- Partial corroboration: OpenRouter's notice independently confirms URL access 40–130 hours after requests to MiniMax models (Hacker News).
Read the full section
- Three breached providers, front-door copying, one shared adversary: secrets from two of the providers were used from the same rented server (Ars Technica). The providers are unnamed, so this rests only on the researcher's report.
- Chinese lab scraping: URLs fetched from more than half of prompts, including about two-thirds of ZDR or "no data collection" requests, typically 2–3 days later (Ars Technica).
- Partial corroboration: OpenRouter's notice independently confirms URL access 40–130 hours after requests to MiniMax models (Hacker News). Dragos does not name the lab, so treating it as MiniMax is an inference.
- A secret link appeared in a Feishu chat 24 days after a ZDR request (Dragos).
Context and prior work
- Days earlier, the Cloud Security Alliance analysed a breach at Fireworks AI, an inference provider used by Notion.
- CSA ties stolen cloud credentials to "LLMJacking," where attackers resell or abuse unauthorised model access. That matches the key abuse seen here.
- OpenRouter's data processing agreement says its ZDR list reflects each provider's stated policy and is not a guarantee that the provider complies (OpenRouter DPA).
Read the full section
- Days earlier, the Cloud Security Alliance analysed a breach at Fireworks AI, an inference provider used by Notion. It exposed AI chat transcripts from under 1% of Notion workspaces, and data persisted outside the ZDR-scoped path (CSA).
- CSA ties stolen cloud credentials to "LLMJacking," where attackers resell or abuse unauthorised model access. That matches the key abuse seen here.
- OpenRouter's data processing agreement says its ZDR list reflects each provider's stated policy and is not a guarantee that the provider complies (OpenRouter DPA).
Limitations, safety and contested findings
- Disputed responsibility: OpenRouter says the problem lies with MiniMax and the hosts that forwarded to it, not with OpenRouter (Hacker News).
- Unidentified parties: the three breached providers and the logging resellers are not named.
- Ongoing abuse: Dragos says credential abuse continued months after it notified providers, and responses were mixed (Dragos).
Read the full section
- Disputed responsibility: OpenRouter says the problem lies with MiniMax and the hosts that forwarded to it, not with OpenRouter (Hacker News). MiniMax attributes the retention to content moderation (TheCyberSecGuru).
- Unidentified parties: the three breached providers and the logging resellers are not named.
- Ongoing abuse: Dragos says credential abuse continued months after it notified providers, and responses were mixed (Dragos).
- Scope of privacy flags: a privacy flag binds only the first hop in a routing chain.
Business and practitioner implications
- Moore's ranking of options: self-host sensitive inference first. Failing that, contract directly with a first-tier provider.
- Prompt hygiene: treat every prompt as readable by each hop. Scrub credentials before sending, and seed your own canaries in agent contexts (Dragos).
- Immediate action: rotate any secrets sent to MiniMax models between June and October 22 (StartupHub.ai).
Read the full section
- Moore's ranking of options: self-host sensitive inference first. Failing that, contract directly with a first-tier provider. Pin upstream providers yourself rather than accepting default router pools (Ars Technica).
- Prompt hygiene: treat every prompt as readable by each hop. Scrub credentials before sending, and seed your own canaries in agent contexts (Dragos).
- Immediate action: rotate any secrets sent to MiniMax models between June and October 22 (StartupHub.ai).
- Procurement: ask vendors to list their inference subprocessors by feature (CSA).
The source trail.
Sources (8)
Researcher plants 2m canaries into the AI supply chain. A privacy disaster surfaced.
Article text retrieved; extracted text may omit tables or interactive elements.
arstechnica.com