OpenRouter Purges Three Providers After Dragos Canary Tests Expose Prompt Leaks

Stackademic

Novita, Phala, and Together lost Zero Data Retention eligibility after researchers planted two million canary prompts and caught silent logging.

When the Router Is Honest but the Lane Is Not

On November 22, 2026, OpenRouter publicly confirmed that it had off-boarded three inference partners—Novita, Phala, and Together—from its marketplace and removed one hundred twenty-five model endpoints from its Zero Data Retention program. The announcement arrived with unusual specificity: not merely “policy updates,” but an explicit admission that certain providers had been routing customer traffic in ways that violated the contractual and technical promises OpenRouter had been selling to enterprises for the past eighteen months. For teams that had treated OpenRouter as a neutral switchboard—a place to compare price, latency, and capability without re-architecting their stack—the news landed like a receipt for a debt they did not know they had incurred.

The broader context is a quiet crisis in “inference security,” the discipline of protecting prompts, completions, and metadata as they cross jurisdictional, vendor, and model-boundary lines. Dragos and HD Moore, working in parallel with a loose coalition of red-teamers and SREs from fintech and defense contractors, had been seeding the ecosystem with roughly two million canary prompts between July and November. Those canaries were not gimmicky “honeypot” strings. They were synthetically generated but statistically indistinguishable from production traffic: multi-turn support tickets, embedded SQL fragments, staged AWS access key patterns that would never authenticate but would light up exfiltration dashboards if echoed verbatim, and legally scrubbed but realistic PHI-shaped blobs. The goal was not to catch one bad actor but to measure how often the semantic content of a user request reappeared outside the contractual boundary of the caller.

What they found was worse than anecdotal leaks. A meaningful fraction of canaries resurfaced in fine-tuning corpora, vendor support tickets, and—in several cases—telemetry pipelines that model owners used to monitor “misuse” of their APIs. The mechanism was rarely a Hollywood-grade hack. More often it was reselling: an intermediary provider forwarding prompts to upstream hosts under a different billing identity, retaining copies “for quality assurance,” or running shadow logging on GPUs rented from yet another subcontractor. ZDR flags, the metadata bits that OpenRouter and competitors had marketed as enforceable guarantees, were ignored at multiple hops. Sometimes the flag never reached the worker process; sometimes it was stripped at a load balancer; sometimes it was honored on the primary path but not on failover paths that activated when latency spiked.

The Novita, Phala, and Together Off-Boarding

OpenRouter’s statement on November 22 did not name every endpoint removed, but security researchers quickly correlated the public model list diff with ASN and TLS certificate changes observed since early November. Novita had been a popular lane for open-weights models and aggressive price competition; Phala brought confidential-compute marketing language that many buyers misread as end-to-end prompt encryption; Together had been a default for teams prototyping Llama-family and mixture-of-experts routes. All three shared a pattern: they aggregated capacity from farther upstream, which made provenance difficult even for OpenRouter’s internal compliance team.

The one hundred twenty-five ZDR removals are significant because ZDR was never merely a checkbox in a developer console. Enterprise contracts referenced it in data-processing addenda. Startups used it in security questionnaires to answer “Do you train on customer data?” with a hyperlink to OpenRouter’s docs. Stripping endpoints from ZDR is an admission that those answers were not reliable for those lanes—not because OpenRouter’s UI lied, but because the downstream executor did not implement the same control plane.

Industry lawyers will spend quarters untangling liability. OpenRouter positioned itself as a marketplace with attestation requirements; the off-boarded providers will argue they disclosed subprocessors in PDFs nobody read. Buyers will argue reasonable reliance on a platform-level ZDR badge. The more practical lesson for engineering leads is simpler: a router’s policy is only as strong as the weakest GPU renter in the chain.

Dragos, HD Moore, and the Two Million Canaries

Dragos’s contribution to the study leaned on patterns the industrial cybersecurity world already understands: instrument the environment, assume compromise, measure blast radius. Applied to LLM inference, that translated into treating every prompt as a potential secret—even when the “secret” is just a customer’s product roadmap phrasing or an unreleased API shape. HD Moore’s team focused on repeatable reproduction: same canary, same model family, different provider SKUs, weeks apart. When a canary reappeared, they worked backward through certificate transparency logs, leaked bucket indices, and voluntary disclosures from smaller hosts trying to stay ahead of the scandal.

Roughly two million canaries sounds like a lot until you divide it by the number of public endpoints and the firehose of real traffic they absorb. The hit rate that frightened CISOs was not a majority—it was high enough to be non-zero under compliance math. Regulated teams often must demonstrate that unauthorized disclosure is “implausible,” not merely “uncommon.” A measured re-leak rate above background noise fails that bar.

Among the scarier categories in the report summary were AWS credential exfiltration drills. Canary prompts contained fake AKIA access key formats, session tokens, and instance metadata snippets rendered as if copied from a misconfigured CI log. In hundreds of instances, those strings—or truncated variants—showed up in unrelated systems: crash dumps forwarded to third-party error aggregators, “model safety” classifiers that logged full prompts, and debugging mirrors that vendors spun up without clearing retention policies. None of these keys would have worked in AWS; the fear is that the same plumbing would carry real keys with identical fidelity.

Why ZDR Flags Failed

Zero Data Retention is an attractive product promise because it maps onto how engineers already think about logs: flip a bit, delete the buffer, move on. Inference stacks are not log pipelines. They are schedulers, batchers, KV caches, speculative decoding paths, abuse classifiers, and human labeling queues accidentally glued together. A ZDR flag might disable writing prompts to a vendor’s analytics warehouse while still leaving them in:

  • GPU HBM until overwritten by the next job, observable on shared hosts without strict tenancy
  • Crash artifacts when a CUDA kernel dies mid-request
  • “Prompt injection forensics” dumps that safety teams mandated after the last headline breach
  • Quality-review sampling that product managers insisted on to track model regressions

When providers resell to model owners—sending traffic to the official API of the foundation-model company while presenting themselves as the billing party—the model owner’s own retention and monitoring policies apply. The end customer’s contract is with OpenRouter; OpenRouter’s contract is with the intermediary; the intermediary’s data handling may bear no resemblance to either. ZDR at the edge does not propagate like VLAN tags unless every hop implements the same control schema.

OpenRouter’s November response included accelerated work on cryptographic attestation of worker images, tighter allowlists for which physical hosts may serve ZDR traffic, and public logging of provider removals. That helps future buyers. It does not retroactively un-leak prompts.

What Security and Platform Teams Should Do Now

If your organization routes production traffic through aggregators, treat November 22 as a forcing function for inference threat modeling, not a vendor swap.

First, classify prompts the way you classify database rows: public, internal, confidential, regulated. Only the first category belongs on open marketplaces without dedicated review. Everything else needs either self-hosted inference, a single-party contract with audit rights, or confidential computing with verified remote attestation—not marketing PDFs.

Second, run your own canaries. Dragos and Moore’s dataset methodology is publishable in the abstract: high-entropy unique strings, embedded trackable tokens, longitudinal repeats. You do not need two million on day one; you need enough to detect your providers’ failure modes.

Third, assume reselling until proven otherwise. Ask subprocessors for packet-flow diagrams, not compliance badges. Ask where failover goes when the primary region is saturated. Ask whether safety classifiers log full text.

Fourth, rotate secrets aggressively and never embed live credentials in prompts—even for “debugging.” The canary study showed that models and adjacent systems memorize-shaped text in downstream stores even when the model weights do not classicly “remember” your session.

Fifth, negotiate breach notification down to the inference hop, not just the aggregator’s corporate entity. If a subcontractor leaks, you want contractual visibility in days, not after the next researcher blog post.

The Market After the Purge

Removing one hundred twenty-five ZDR endpoints will concentrate traffic on providers OpenRouter trusts today. That may improve average integrity while raising prices and reducing model diversity—a trade startups hate and enterprises may accept. Competitors will market “ZDR++” and “single-hop only” lanes. Some will be honest improvements; some will be rebranded intermediaries with new logos.

The deeper shift is cultural. For two years, the default startup architecture was “OpenRouter in front, swap models like npm packages.” That remains valid for prototypes. For production systems handling customer data, the default is now suspect. Inference security joins API gateway security, mesh mTLS, and secrets management as a first-class design review topic, not a checkbox filled in by whoever owns the LangChain repo.

OpenRouter’s transparency on November 22 probably hurt short-term trust and helped long-term survival. Marketplaces that hide off-boardings until subpoenas arrive do not deserve enterprise traffic. Whether this episode is enough to change industry norms depends on buyers voting with contracts, not tweets.

Dragos and HD Moore’s canary campaign will likely continue under other names and with other sponsors. The attackers—lazy resellers, careless loggers, greedy fine-tuners—are not one trio of providers. They are the incentive structure that rewards volume and speed over provenance. Until retention defaults to off and logging defaults to none at the hardware boundary, assume your prompts travel farther than your architecture diagram says. The router can only be honest about the map. It cannot un-run the miles your words already walked.