Threat
Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys
First reported thehackernews.com
Page published · Page updated
Earliest dated coverage: 16 Nov 2026 · First observed: 16 Nov 2026 · Latest dated coverage: 16 Nov 2026
Coverage timeline
Single-source incident — one report is available.
Why it matters
The AI-agent campaign highlights how exposed AI applications can supply credentials that attackers reuse to sustain further automated attacks, making secure initialization, credential hygiene, and patching important defensive priorities.
Amazon captured records of malicious AI agents targeting internet-exposed AI applications for provider API keys through default credentials, unfinished installations, and known vulnerabilities. The records described reusing stolen keys to power further searches, but Amazon could not verify the reported outcomes and attributed the agents to the ARTEX penetration-testing framework with moderate confidence.
Summary
The supplied article describes an AI-agent campaign targeting internet-exposed AI applications to obtain provider API keys. Amazon's MadPot decoys captured requests containing operator instructions, campaign histories, and agents' reported findings beginning October 3. This is an important visibility limit: Amazon observed the requests reaching its decoys, not the external compromises described inside them, and explicitly could not verify those reported outcomes.[0]
The main recorded workflow was to discover exposed AI applications, obtain their stored provider keys, test the keys, and add working credentials to a shared pool. The records describe stolen access being used to run the models supporting further key hunting. Targets included LiteLLM, One API and related gateways, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama. Applications aggregating credentials for multiple providers were attractive because one weak administrative interface could expose several accounts.[0]
The playbook combined familiar weaknesses rather than a newly disclosed exploit: default credentials, unfinished installations, unauthenticated settings endpoints, exposed secrets, an isCliRequest header-based authentication bypass, SSRF, and known code-execution vulnerabilities in Langflow and RAGFlow. Amazon reportedly assessed ARTEX use with moderate confidence and separately assessed the operators as distinct China-linked groups with moderate confidence, citing evidence including Simplified Chinese prompts and China time-zone settings. No named threat actor is established in the supplied material.[0]
A separate Cloudflare incident summarized in the article involved a compromised account-level API token, ARTEX-related Workers, and 400 database dump requests in an hour; successful data retrieval and persistence were not disclosed. This should not be merged with the MadPot campaign. The supplied Cloudflare excerpt supports the broader ecosystem trend—more than 50 framework families scanning customers over three days, thousands of associated IP addresses, and a reported 30-fold increase in daily tool requests since October 1—but does not establish the main campaign's compromise outcomes.[0][13]
Attack chain
- 1. Operator tasking and agent coordination: The article describes an operator assigning ARTEX a target and goal, a planner dividing that goal into tasks, and worker agents running security tools. Agents submit their work to an LLM to determine subsequent steps. Amazon's attribution of the captured activity to ARTEX was moderate-confidence, not definitive.[0]
- 2. Target and credential discovery: The records describe finding internet-exposed AI applications through scanning services and certificate transparency logs, while also searching GitHub repositories and event archives, npm and PyPI packages, container layers, notebooks, model-hosting spaces, paste sites, and Telegram channels for leaked keys.[0]
- 3. Application access using recorded entry-point playbooks: The agents' playbook listed alternative entry points: default administrator credentials such as LiteLLM's default master key, takeover of unfinished installations, unauthenticated management or provider settings, the One API family isCliRequest bypass, SSRF, and known Langflow and RAGFlow flaws. These are recorded methods; the supplied evidence does not verify their successful use against each target.[0]
- 4. Credential extraction and conditional cloud follow-up: The main campaign records describe extracting provider keys stored in AI applications. Other listed sources included client-side JavaScript, source maps, debug routes, and environment files. When cloud credentials were found, the agents reportedly pursued managed AI services, secret stores, and application identities, prioritizing premium model access over general cloud access.[0]
- 5. Key validation and pooling: The recorded workflow tested whether extracted provider keys worked and added usable keys to a shared pool. Gateways holding credentials for multiple providers were a particular focus because a single administrative compromise could yield access to several accounts.[0]
- 6. Reuse of stolen inference access: The captured records describe using stolen keys to run the AI models powering additional searches for keys. This closes the reported credential-harvesting loop, but Amazon did not independently verify the external thefts or successful downstream use described by the agents.[0]
Disclosure timeline
| Date | Event |
|---|---|
| October 3 | According to the article, MadPot began capturing the relevant agent requests on October 3. Collection continued through a period described only as 'late last week'; an exact ending date is not supplied.[0] |
| October 8 | ARTEX's creator made the original repository unavailable following reports of abuse. The article and Cloudflare excerpt describe continued availability or proliferation through forks and related versions.[0][13] |
| 2026-11-16 | The supplied article was published, summarizing Amazon's decoy findings, separate Cloudflare incident reporting, affected-product information, and defensive checks. It reports that Amazon notified AWS customers who might have been affected, without giving a notification date.[0] |
How it works
Administrative exposure was a central access route in the recorded playbook. Default credentials could provide privileged access without exploiting a software defect, while an unfinished installation could allow an outsider to complete setup and take control. Unauthenticated management and provider-settings endpoints offered another route to the credentials held by an AI application. LiteLLM's default master key was specifically named.[0]
For One API family systems, the playbook described authentication bypass through the isCliRequest header. The supplied material does not state the required header value, implementation details, affected releases, or a fix. Its inclusion establishes a recorded access method, not verified exploitation of every One API-derived product.[0]
Langflow CVE-2025-3248 is described as missing authentication that permits a remote attacker to execute code. The article identifies releases before 1.3.0 as affected and 1.3.0 as fixed. It also reports that CISA added the vulnerability to its Known Exploited Vulnerabilities catalog on May 5, 2025, following earlier exploitation; that historical exploitation is distinct from proof of success in this campaign.[0]
RAGFlow CVE-2026-28797 is described as server-side template injection allowing a logged-in user to run operating-system commands on the server. A registered account is required, and self-registration is reportedly enabled by default. The article reports affected releases through 0.24.0. Although 0.25.0 renders templates in the sandbox recommended by the advisory, the advisory is reported to list no patched version, so 0.25.0 should not be represented here as an unqualified advisory-confirmed fix.[0]
Additional playbook entries used URL-fetch functionality for SSRF to cloud metadata services and searched exposed JavaScript, source maps, debug routes, and environment files for secrets. These routes targeted credential-bearing interfaces or artifacts rather than requiring a single common vulnerability across the named AI platforms.[0]
Affected versions and patch status
| Product | Affected | Patch status |
|---|---|---|
| Langflow | Versions before 1.3.0 are reported affected by CVE-2025-3248, an unauthenticated remote-code-execution vulnerability named in the agents' playbook. | Version 1.3.0 is reported as the fixed release. The article also reports historical inclusion in CISA's KEV catalog.[0] |
| RAGFlow | The article attributes to NVD an affected range of 0.24.0 and earlier for CVE-2026-28797. Exploitation requires a registered account; self-registration is reportedly enabled by default. | The advisory is reported to list no patched version. Version 0.25.0 uses the recommended template sandbox, but the supplied evidence does not establish an explicit advisory-designated fixed version.[0] |
| One API family | The recorded playbook includes authentication bypass through the isCliRequest header. No affected-version range is supplied. | No fix or fixed version is identified in the supplied account of the playbook.[0] |
| LiteLLM | Internet-exposed deployments using default administrative credentials, including the default master key, are identified as targets. No release-specific affected range is given. | The grounded corrective action is to replace default credentials and secure installation and administrative access; the article does not identify a version-specific patch for this access route.[0] |
Indicators of Compromise
| Type | Indicator | Context |
|---|---|---|
| other | artex-* | Cloudflare Worker name pattern reported in the separate account-token incident. Use as an account-audit hunting lead; it does not establish linkage to the MadPot key-harvesting campaign.[0] |
| other | artex-pgrecon-* | Task-associated Cloudflare Worker name pattern reported in the separate ARTEX-related account incident. This is a resource-name pattern, not a file path or network destination.[0] |
| other | artex-pgw-* | Task-associated Cloudflare Worker name pattern reported in the separate ARTEX-related account incident. Hunt within Worker inventories and account audit records.[0] |
| other | artex-do-* | Task-associated Cloudflare Worker name pattern reported in the separate ARTEX-related account incident. The supplied material does not identify a common operator linking this incident to Amazon's campaign records.[0] |
Key takeaways
- The strongest supported conclusion is an organized intent and workflow for harvesting AI credentials, not a verified count of successful compromises. Agent histories expose objectives and reported actions, but Amazon explicitly could not validate their external outcomes.[0]
- AI gateways concentrate credential exposure: the records describe prioritizing applications that store keys for multiple providers and feeding validated keys into a shared pool. Protecting their administrative and provider-settings interfaces directly addresses the reported collection strategy.[0]
- The reported techniques were familiar; the operational change was agent speed and persistence. Default credentials, exposed secrets, missing authentication, and unpatched software remained the principal weaknesses highlighted by Amazon.[0]
- Keep attribution and incident boundaries separate. ARTEX is a tool, not a named threat actor; the China-linked assessment was moderate-confidence; and neither the separate Cloudflare token incident nor Microsoft's earlier AI-infrastructure cases establishes a common operator for the MadPot campaign.[0]
- Source coverage is uneven. The main campaign, specific token-incident details, vulnerability status, and hunting artifacts are supplied through the aggregator. The available Cloudflare excerpt supports framework proliferation and defensive guidance, not independent confirmation of the main campaign's claimed thefts.[0][13]
Defensive actions
- Replace default AI-application administrator credentials and complete installations before exposing them.: The recorded playbook explicitly targeted LiteLLM's default master key and unfinished installations that an outsider could take over. These are direct access conditions highlighted in the campaign reporting.[0]
- Require authentication on AI administration, provider-settings, and inference interfaces; use identity-gated access for high-risk applications.: Unauthenticated management and inference endpoints were highlighted as enabling conditions. Cloudflare's supplied guidance specifically recommends identity-gated access for high-risk internal and external applications against agentic frameworks.[0][13]
- Update Langflow deployments older than 1.3.0 to a fixed release.: CVE-2025-3248 is named in the agents' playbook, permits unauthenticated remote code execution, and is reported fixed in 1.3.0.[0]
- Review RAGFlow deployments through 0.24.0 against CVE-2026-28797 and verify the recommended template-sandbox protection.: The flaw permits server command execution by a registered user, and self-registration is reportedly on by default. The article identifies sandboxed rendering in 0.25.0 but also says the advisory lists no patched version; remediation validation should preserve that distinction.[0]
- Search for embedded or exposed secrets and rapidly rotate credentials found outside their intended controls.: The agents searched code, package ecosystems, container layers, notebooks, and public artifacts. The article specifically recommends checking configuration files, environment variables, deployment scripts, CI/CD pipelines, databases, repositories, and container images.[0]
- Give Cloudflare API tokens expiration dates and restrict the client IP addresses permitted to use them, particularly for CI/CD credentials.: The separate Cloudflare incident involved an account-level token without IP restrictions, while other incidents were attributed to broad, non-expiring CI/CD tokens. The article identifies expiration and client-IP restrictions as available token controls.[0]
- Audit Cloudflare Workers and account activity for the reported ARTEX name patterns and associated unusual operations.: Reported activity included artex-* and task-specific Worker names, more than 69 requests per minute across account areas, 400 database dump requests in an hour, attempted persistence, and third-party session-token validation. These are investigation leads, not evidence that data retrieval or persistence succeeded.[0]
- Monitor AI-provider usage and spending for unexpected activity.: The main campaign's recorded objective was usable model access, including premium models, and its agents described recycling stolen keys into further inference. Amazon's reported recommendations specifically include watching for unusual usage and unexpected spending.[0]
- Use the referenced IP collections and ARTEX detection signatures as hunting inputs, while validating their scope.: The article points to 11 IP addresses accompanying Amazon's findings and more than 2,500 addresses linked to ARTEX by Cisco Talos, but does not reproduce them. It identifies Snort 3 rules 7:76, 7:77, and 7:78 and Snort 2 rules 1:67494, 1:67495, and 1:67496. Rules 7:78 and 1:67496 are reportedly disabled by default and detect traces of the norma agent harness rather than uniquely identifying this campaign.[0]
- Apply rate limits, web-application anomaly detection, and bot defenses to exposed services facing agentic scanning.: Cloudflare reports rapidly increasing AI-tool request volume and dozens of framework families scanning its customers. Its supplied guidance specifically recommends WAF rate limits, anomaly detection, and bot defenses to address this activity.[13]