abilities

A sweeping industrial cybersecurity investigation has revealed that malicious actors secretly intercepted data from inside multiple artificial intelligence inference providers for months without detection. According to findings published by industrial cybersecurity firm Dragos, none of the affected providers were aware of the continuous data exfiltration until independent researchers alerted them. The breaches heavily compromised sensitive prompts and infrastructure, casting a harsh spotlight on the systemic vulnerabilities embedded within the modern artificial intelligence supply chain—particularly concerning settings marketed as strict privacy guarantees.

Among the intercepted communications were prompts explicitly flagged as "Zero Data Retention," a widely utilized configuration intended to ensure that user inputs are never stored, logged, or retained by processing infrastructure. The revelations underscore a profound security gap in how artificial intelligence routing works behind the scenes, raising urgent questions about data handling practices across third-party cloud platforms, API marketplaces, and decentralized model resellers.

The Scope of the Breaches and Industry Response

The widespread security failures came to light following an extensive research initiative conducted between July and November. Using sophisticated canary token research, researchers seeded approximately two million distinct AI prompts with unique, untraceable web links and credentials. By continuously monitoring for unauthorized attempts to access these planted secrets, the cybersecurity team could precisely track when and where prompts were being read, copied, or diverted. Because a planted credential can only be discovered if an external entity actually reads the prompt carrying it, the method provided undeniable proof of widespread data leakage.

OpenRouter, a prominent gateway that dynamically routes customer artificial intelligence requests to a diverse pool of third-party inference providers, responded swiftly to the disclosures. The company announced that it has removed three distinct providers from its platform, stripped 125 specific endpoints of their eligibility for Zero Data Retention, and directly notified all potentially affected customers. Notably, OpenRouter emphasized that its own core infrastructure experienced no direct security breach, attributing the failures entirely to downstream inference partners.

The earliest of the identified breaches traces back to mid-July. Although the prompts planted by researchers were specifically designed as security bait, they were processed concurrently alongside authentic, real-world customer requests. Researchers confirmed that genuine customer traffic was almost certainly copied and exfiltrated alongside the canary tokens. The largest of the documented incidents involved several hundred Amazon Web Services (AWS) keys that were planted in the context windows, each dispatched in precisely one request to a single provider during brief operational windows in late October.

Every single request transmitted during those windows was successfully stolen. Alarmingly, some of the compromised data originated from requests that the inference provider had formally rejected before an AI model even had a chance to execute. This critical detail indicated that attackers were copying network traffic at the absolute earliest point of arrival—at the perimeter of the infrastructure stack—rather than deep within the model execution layers.

Sophisticated Attack Vectors and Infiltration Techniques

The attackers wasted little time weaponizing the stolen data. In several instances, malicious actors tested newly acquired AWS keys within hours of theft. They systematically listed storage buckets, scoured Secrets Manager and Parameter Store configurations, and attempted to generate unauthorized Bedrock API keys. Furthermore, they tried to create new Identity and Access Management (IAM) users to establish persistent, long-term access, driving tens of thousands of automated AWS API calls in the process.

OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms

In one notable convergence of malicious activity, a single threat actor appeared to compromise two distinct inference providers simultaneously, utilizing stolen secrets harvested from both targets from a single rented server. By November 1, this same actor integrated ARTEX, an advanced artificial intelligence-driven penetration testing framework, using it exclusively to verify the ongoing validity of the keys they already possessed. Researchers noted that the operation utilized specific user agents—namely artex-agent and artex-judge-r121—that did not appear on any public indicator of compromise lists, making them critical additions for defensive detection engineering.

"A privacy flag doesn’t help when the attacker is already inside," remarked HD Moore, chief research officer at Dragos, encapsulating the fundamental flaw in relying solely on logical privacy settings when underlying server infrastructure is compromised.

Beyond direct server compromises, the research uncovered security risks that required no active break-in at all. Certain sellers operating within AI inference marketplaces deliberately listed advanced models at artificially low prices, explicitly enabled request logging, and passively harvested the valuable data flowing through their systems. The individuals operating these collection nodes often appeared to be casual users logging in from residential internet connections. For instance, one operator working from a Mac on home broadband in France successfully extracted passwords from intercepted prompts and manually attempted root Secure Shell logins, before using the exact same machine to test planted AWS credentials.

Further complicating the threat landscape, network routing anomalies played a role in data diversion. One tested credential had traveled through a third-party router that quietly and without customer knowledge forwarded a portion of its traffic directly to an inference marketplace. That network routing entity later confirmed to researchers that it had ceased the practice. However, subsequent route tracing across approximately 2,000 distinct paths revealed several hundred unidentified networking systems, alongside one specific marketplace seller that surreptitiously injects roughly 1,000 hidden tokens into every customer prompt that passes through its infrastructure.

International data leakage was also documented during the months-long study. An unnamed Chinese artificial intelligence laboratory fetched secret tracking links from more than half of the seeded prompts sent to its proprietary models between July and October. This included roughly two-thirds of all requests that carried strict no-data-collection or Zero Data Retention designations. Prompts dispatched to United States-based hosting providers that merely resell the foreign lab’s models triggered the exact same automated scraping behavior, causing sensitive data to flow directly back to the lab.

In another instance, a Git client operating on a separate cloud platform later attempted to clone private code repositories referenced within the intercepted prompts, though strictly for that specific lab’s models. Meanwhile, a large U.S.-based hosting provider continuously crawled a secret canary link for more than two months following a single initial request to its search endpoint. Twenty-four days after one specific test prompt was sent, its secret link surfaced inside a chat group on Feishu, the Chinese workplace collaboration application known internationally as Lark. The link was accessed within two minutes from three separate desktop computers and three mobile devices located in Beijing and Shanghai. Because the researchers had publicly disclosed only the initial eight characters of the canary link, whoever posted it was clearly reading a fully cached, unredacted copy of the original prompt.

What the Providers Have Disclosed

In the wake of the coordinated disclosures, several prominent inference providers have stepped forward with public incident reports detailing how their environments were compromised.

Phala disclosed in a preliminary incident report that an unauthorized attacker successfully seized control of an internal inference server, granting them visual access to request data while it was actively being processed. The company stated that the security event was successfully contained on November 8, and normal model services have since been restored alongside reinforced security architectures. Phala advised that users who submitted requests to the DeepSeek V4 Flash 0731 model via the OpenRouter gateway between October 21 and November 8 should treat their data as fully exposed and immediately rotate all associated keys, tokens, and passwords. While an independent researcher had placed canary credentials into requests starting July 1, only the prompts transmitted on October 21 and 22 were actively abused by the intruder. Because the testing initiative only sampled a fraction of total traffic rather than monitoring every single request, Phala maintained a broader precautionary timeline for its advisory.

OpenRouter Drops 3 Providers After Researchers Catch Months of Prompt Theft at AI Hosting Firms

Together AI issued a security notice revealing that an unauthorized party managed to download stored samples of customer prompts. The breach occurred after an artificial intelligence coding agent inadvertently committed a sensitive API credential to the company’s public GitHub repository. The compromised samples had been systematically collected for routine software engineering testing and originated from customer requests directed toward MiniMax M3, GLM-5.2, and Kimi K2.6 models during June and July. The malicious actor searched the harvested samples for credentials belonging to external services and attempted to validate them, ultimately encountering five canary tokens planted by the researchers. Together AI stated it has found no evidence that the exfiltrated data spread further, believes the files have been permanently deleted, and has subsequently transitioned Zero Data Retention to the default setting for all newly registered user accounts.

Novita Labs confirmed through a privacy notice that an unintended routing configuration error resulted in certain MiniMax inference requests being inadvertently forwarded directly to the official MiniMax API, exceeding the intended processing pathways for those users.

Neither Dragos nor OpenRouter publicly identified the exact identities of all three originally breached providers, and it remains unconfirmed whether the entities named in public disclosures overlap entirely with the gateway’s removed partners. However, researchers noted that the organization responsible for the mid-July breach initially attempted to blame internal misconfigurations, subsequently pointed toward a partner’s data processing pipeline, and only realized the truth after internal forensic investigations uncovered compromised credentials. That specific provider had still not issued a public disclosure regarding the incident as the findings were made public.

Recommendations for Securing AI Pipelines

As the industry grapples with the far-reaching implications of these supply chain vulnerabilities, cybersecurity experts emphasize that traditional logical boundaries are insufficient when dealing with compromised infrastructure. HD Moore noted that a privacy flag is fundamentally constrained to the very first hop of a transaction, and autonomous AI agents will inevitably transmit whatever readable data they are given, including critical system credentials. Where an enterprise prompt travels remains a deliberate architectural choice; organizations that fail to make deliberate routing choices effectively delegate that security decision to the lowest-cost provider in the marketplace.

For organizations handling highly sensitive intellectual property, self-hosting remains the only definitive method to ensure that no third-party entity reads user prompts. When utilizing commercial services, routing directly through established, first-tier providers helps minimize administrative complexity, establishing a single contract and a single privacy policy to evaluate.

For those utilizing API routers and multi-model gateways, experts advise exercising heightened caution. Organizations should manually select upstream providers rather than accepting default routing pools, rigorously verify whether privacy flags genuinely alter physical request pathways, and entirely avoid unverified reseller marketplaces whose underlying capacity may rely on stolen infrastructure or compromised keys.

Implementing internal tripwires—such as embedding free canary tokens within context windows routinely processed by software agents and automated pipelines—can provide immediate alerts if sensitive secrets are exfiltrated. While automated filters designed to scrub common credential formats from prompts offer a baseline layer of defense, they frequently fail to capture custom secrets or sensitive corporate data that does not conform to standard key structures.