{"schema":"sanctity-signal/0.2","record_type":"signal","id":"dragos-ai-prompt-canaries-e108","domain":"security","drop_id":"signal-e108","occurred_at":"2026-11-19","how_to_use":"This is a structured record of one real-world action and its competing interpretations, from Sanctity Signal. To build content from it (a reel script, a blog, a thread): open on the hook, ground the viewer with plain and observed, walk the dated trail, present the interpretations as live tension between named readings, and close on the before/after shift and what_to_watch. Hard rules: items under established are fact; items under unknown are never presented as fact; the evidence status (verified or reported) is stated, not dropped; sources are cited.","actor":"Dragos","action_type":"publication of a 5 month experiment that seeded nearly 2 million AI prompts with unique credentials and URLs, traced where they travelled, and found active theft at 3 hosting providers","observed":"Dragos published AI Supply Chain Security, What Canary Tokens Found, by chief research officer HD Moore, on November 19, 2026. The page header gives 08:00; the metadata gives 13:00 with no time zone. Its core claim: Dragos sent nearly 2 million AI prompts containing unique, trackable secrets to AI providers, over 5 months from July to November 2026. Each prompt carried unique URLs and canary credentials, including AWS keys and DNS names. The work began with Thinkst Canarytokens in July, then Dragos built a seeding service that derives unique hostnames, URLs and credentials per prompt and hosts the services they point at, covering DNS, web, git, SSH, databases and mail. AWS keys were tied to empty, fully logged accounts. Routes were traced with image URL callbacks and malformed requests that expose nested error envelopes from each hop; about 2,000 routes were traced and several hundred included systems Dragos could not identify. Findings. Attackers actively stole data from 3 hosting providers, the earliest incident dating to mid-July. The largest breach involved a couple hundred AWS keys, each sent in a single request to one provider during late-October windows; every request in those windows was stolen, and some were taken before a model ran, which points at copying at the provider's front door. Two of the breaches included requests marked zero data retention. Dragos writes that a privacy flag does not help when the attacker is already inside. Stolen keys were used for tens of thousands of AWS API calls, and one adversary hit 2 providers at once from a single rented server. On November 1 the same attacker used ARTEX, a model-driven penetration testing framework, sending the user agents artex-agent and artex-judge-r121. Providers were unaware until notified; one has still not disclosed publicly, another published a preliminary notice. Resellers obtained prompts with no exploit at all, through ordinary request logging: Dragos writes that sellers only had to list a model at the lowest price, enable request logging, and wait. Secret URLs were opened about a day and a half after a request, followed by login guessing, including an SSH root attempt using passwords taken from prompts. A Chinese AI lab fetched secret URLs from more than half of prompts sent between July and October, including roughly two-thirds of requests marked no data collection or zero data retention, with scrapers typically appearing 2 to 3 days later.","corrections":[],"evidence":{"status":"verified","note":null,"sources":[{"role":"primary","label":"Dragos: AI Supply Chain Security, What Canary Tokens Found, by HD Moore, Nov 19, 2026","url":"https://www.dragos.com/blog/ai-supply-chain-security-canary-tokens"},{"role":"corrob","label":"HD Moore on X: 2 million canaries seeded, 3 compromised providers, resellers stealing prompts via logging","url":"https://x.com/hdmoore/status/2123451306685661416"},{"role":"corrob","label":"Cybrsecmedia: Two million AI prompts expose breaches, data theft and supply chain failures","url":"https://www.cybrsecmedia.com/two-million-ai-prompts-expose-breaches-data-theft-and-supply-chain-failures/"}]},"strength":{"cost":"High","commitment":"High","reversibility":"Hard","persistence":"Structural","rationale":"Cost is high. Five months, nearly 2 million prompts, a purpose-built seeding service hosting 6 kinds of fake service, 2,000 traced routes and real AWS accounts left open to be abused is a research programme. Commitment is high: a named chief research officer puts his firm behind findings that accuse unnamed providers of being breached, which invites contradiction from anyone who can produce logs. Reversibility is hard, because the evidence is the callbacks, which happened and are logged. Persistence is structural: the finding is not a bug that gets patched but a property of a supply chain where a prompt passes through hops the sender never chose, and the counting method now exists for anyone to run."},"shift":{"prior":"A privacy flag on a prompt kept it out of other hands.","action":"Dragos published AI Supply Chain Security, What Canary Tokens Found, by chief research officer HD Moore, on November 19, 2026. The page header gives 08:00; the metadata gives 13:00 with no time zone. Its core claim: Dragos sent nearly 2 million AI prompts containing unique, trackable secrets to AI providers, over 5 months from July to November 2026. Each prompt carried unique URLs and canary credentials, including AWS keys and DNS names. The work began with Thinkst Canarytokens in July, then Dragos built a seeding service that derives unique hostnames, URLs and credentials per prompt and hosts the services they point at, covering DNS, web, git, SSH, databases and mail. AWS keys were tied to empty, fully logged accounts. Routes were traced with image URL callbacks and malformed requests that expose nested error envelopes from each hop; about 2,000 routes were traced and several hundred included systems Dragos could not identify. Findings. Attackers actively stole data from 3 hosting providers, the earliest incident dating to mid-July. The largest breach involved a couple hundred AWS keys, each sent in a single request to one provider during late-October windows; every request in those windows was stolen, and some were taken before a model ran, which points at copying at the provider's front door. Two of the breaches included requests marked zero data retention. Dragos writes that a privacy flag does not help when the attacker is already inside. Stolen keys were used for tens of thousands of AWS API calls, and one adversary hit 2 providers at once from a single rented server. On November 1 the same attacker used ARTEX, a model-driven penetration testing framework, sending the user agents artex-agent and artex-judge-r121. Providers were unaware until notified; one has still not disclosed publicly, another published a preliminary notice. Resellers obtained prompts with no exploit at all, through ordinary request logging: Dragos writes that sellers only had to list a model at the lowest price, enable request logging, and wait. Secret URLs were opened about a day and a half after a request, followed by login guessing, including an SSH root attempt using passwords taken from prompts. A Chinese AI lab fetched secret URLs from more than half of prompts sent between July and October, including roughly two-thirds of requests marked no data collection or zero data retention, with scrapers typically appearing 2 to 3 days later.","update":"The flag binds the first hop. Attackers and resellers sit past it."},"inference":{"established":["Attackers stole data from 3 AI hosting providers, the earliest from mid-July.","Two breaches included requests marked zero data retention.","Stolen keys were used for tens of thousands of AWS calls."],"unknown":["Which providers. Dragos names none of the compromised hosts.","Who the attackers are. One used a tool called ARTEX.","That a customer lost money. Only Dragos's own planted keys were used."]},"narrative":{"hook":"Dragos seeded 2 million AI prompts with traps. Attackers were inside 3 providers.","plain":"Over 5 months Dragos sent nearly 2 million prompts carrying unique credentials and URLs, then watched who used them. Two of the breached providers were serving requests marked zero data retention.","story":["The method is cheap enough to copy, and it measures the AI supply chain from the inside rather than from a vendor's word. Anyone sending prompts through a router or a marketplace now has a way to find out where they actually go, and a reason to assume nobody has checked."],"trail":[{"d":"JUL","t":"Dragos starts with Thinkst Canarytokens, then builds its own seeding service. The earliest provider compromise it later dates also begins in mid-July."},{"d":"JUL TO OCT","t":"A Chinese AI laboratory fetches secret URLs from more than half of prompts sent in the period, including about two-thirds of those marked no data collection or zero data retention."},{"d":"LATE OCT","t":"The largest breach: a couple hundred planted AWS keys taken from one provider, every request inside the windows stolen, some before a model ran."},{"d":"NOV 1","t":"The same attacker runs ARTEX, a model-driven penetration testing framework, against the planted AWS accounts, identifying itself as artex-agent and artex-judge-r121."},{"d":"NOV 19","t":"Dragos publishes the method, the findings and its recommendations. It began notifying providers in early November and names none of the 3 compromised hosts."},{"d":"Reacts","r":true,"t":"The 3 unnamed providers and the dozen Dragos is working with; OpenRouter and other routers, one of which appears as an undisclosed hop; every enterprise that bought zero data retention as a control; cloud security teams whose keys travelled in prompts; and regulators who have been told prompts are not retained."}],"glossary":[{"term":"Canary token","def":"A fake secret planted so its use gives away a leak. Nobody legitimate would ever touch it, so any use is evidence."},{"term":"Zero data retention","def":"A provider setting that promises a request is not stored. It binds the provider you call, not whoever sits behind it."},{"term":"Inference reseller","def":"A marketplace or router that sells access to someone else's model, passing your prompt along one or more hops."}]},"judgment":{"prompt":"Which explanation do you believe?","interpretations":[{"id":"a","label":"The supply chain is the weakness","text":"Resellers needed no exploit: list cheap, log requests, wait.","abstain":false},{"id":"b","label":"Three unlucky providers","text":"3 hosts, against a dozen now cooperating. The rest showed nothing.","abstain":false},{"id":"c","label":"A deliberate collection operation","text":"Every request in the October windows taken, some before a model ran.","abstain":false},{"id":"d","label":"Ordinary cloud insecurity, new label","text":"Stolen keys and open logging are old problems wearing an AI name.","abstain":false},{"id":"e","label":"Too early to call.","text":"The evidence cannot yet separate these.","abstain":true}],"distribution":{"status":"held back until you have made your call","calls":null}},"story_thread":null,"what_to_watch":[{"linked_interpretation":"a","direction":"strengthens","observable_event":"A named inference reseller or router is shown to have logged or resold customer prompts, or a provider discloses a breach matching the late-October windows, by March 31, 2027.","window_end":"2027-03-31","state":"unresolved"},{"linked_interpretation":"c","direction":"strengthens","observable_event":"Any provider, vendor or government attributes the late-October theft or the ARTEX activity to a named actor, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"},{"linked_interpretation":"a","direction":"weakens","observable_event":"Two or more of the 3 providers publish post-mortems showing the theft came from an ordinary credential compromise unrelated to resale or routing, by June 30, 2027.","window_end":"2027-06-30","state":"unresolved"}]}