Threat
Attackers are using AI agents to automate cybercrime at 'industrial scale,' FBI and AWS warn
First reported theregister.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
ARTEX-enabled credential theft creates a self-reinforcing attack cycle in which compromised model access powers further automated searches for AI credentials.
Amazon Threat Intelligence reports that attackers used the open-source ARTEX AI penetration-testing framework to steal AI API keys and cloud credentials, then use stolen model access to power agents hunting for additional keys. A separate ARTEX campaign attempted cryptocurrency-exchange identity-verification bypasses using deepfakes, although the described vendor's server-side liveness checks rejected every attempt.
Summary
The Register's November 16, 2026 report describes FBI and Amazon Threat Intelligence warnings about criminal use of the open-source ARTEX penetration-testing framework. According to the article, Amazon's MadPot sensors observed two campaigns between October 3 and November 5: one stole AI-access credentials and reused them to sustain further agent activity; another attempted to defeat cryptocurrency-exchange identity checks. The broader warning covered banks, government organizations, energy and healthcare providers, and other critical sectors.[0]
The credential campaign created a reinforcing access-theft workflow: agents stole API keys, supplied them to ARTEX's model gateway, and used the resulting model access to run more agents searching for additional keys. Cloud credentials also supported identity validation, storage enumeration, permission assessment, and pursuit of AI services and secrets. The reported targets included both LLM gateway/proxy products and AI application builders.[0]
The identity-fraud campaign combined parallel work on deepfake video, verification bypass, account balances, and cash-out research. Its demonstrated technical progress did not establish a successful end-to-end bypass: although agents completed one vendor's requested face-movement sequence within seven pixels of the target, the server-side verification decision rejected every reported attempt. A separate session achieved 11.5 frames per second for real-time face animation, while other work sought to make forged ID images retain plausible camera and JPEG characteristics.[0]
The defensive significance is increased attacker speed and persistence rather than a newly identified exploit technique. AWS reportedly observed success where established controls were missing, while the FBI warned that some intrusions reached privileged access in under three hours. All findings here rest on the supplied Register article; the linked AWS and FBI publications were not supplied as source bodies and therefore do not provide independent corroboration.[0]
Attack chain
- Credential campaign — acquire access keys: Agents stole AI API keys. The reported target set included LLM gateways, proxy platforms, and AI application builders; the campaign also used cloud credentials to obtain unauthorized model access.[0]
- Credential campaign — supply the ARTEX model gateway: Attackers fed stolen API keys into ARTEX's model gateway, making the stolen model access available to support additional agent activity.[0]
- Credential campaign — repeat credential discovery: The newly powered agents searched for more AI model-access keys, extending the same credential-theft loop. The report does not quantify the number of keys stolen or the resulting model usage.[0]
- Independent KYC campaign — reconstruct the liveness protocol: Agents reverse-engineered one verification vendor's face-detection event stream, face-position tracking, and movement-coordinate system. This was a separate campaign, not a demonstrated continuation of the credential-theft operation.[0]
- Independent KYC campaign — execute the movement challenge: The agents completed the full requested movement sequence with the generated face positioned within seven pixels of the target. Parallel sessions worked on deepfake video, forged documents, account balances, and possible cash-out methods.[0]
- Independent KYC campaign — server-side rejection: The vendor's server-side verdict rejected the deepfake in every reported attempt. Completing the visible movement challenge therefore did not demonstrate successful identity verification.[0]
Disclosure timeline
| Date | Event |
|---|---|
| 2026-08 | The FBI reportedly cited ARTEX-associated attacks against South Korean financial institutions that stole personal data belonging to approximately 67,000 people. The article provides the month, not an exact incident date.[0] |
| 2026-09 | The FBI reportedly cited an ARTEX campaign against ConnectWise Cloud Backup customers that ultimately compromised more than 2,250 downstream Microsoft tenants.[0] |
| 2026-10 | The original ARTEX developer reportedly removed the GitHub repository after earlier breaches became public. Forked and localized variants and a commercial version remained available.[0] |
| 2026-10-03 to 2026-11-05 | Amazon's MadPot sensors detected the AI-credential-theft and cryptocurrency identity-verification campaigns during this observation window.[0] |
| 2026-11-04 | Censys reportedly identified 1,465 hosts running an exposed ARTEX console. The FBI cautioned that deployments behind firewalls or on private networks would increase the overall deployment count.[0] |
| 2026-11-13 | AWS published its campaign account and the FBI issued a separate alert, identified in the November 16 article as the preceding Friday's publications.[0] |
| 2026-11-16 | The Register published its account combining the AWS observations, FBI warning, and reported exposure findings.[0] |
How it works
The credential operation depended on obtaining usable credentials rather than on a single disclosed software vulnerability. Stolen API keys were inserted into ARTEX's model gateway and used to power agents that searched for further keys. AWS's reported explanation connected successful access to default or hardcoded credentials, keys in code and public artifacts, unauthenticated interfaces or inference endpoints, and unpatched software.[0]
Cloud-access activity included validating identities through provider token and identity services, enumerating storage, assessing permission scope, and pursuing managed AI services, secrets and parameter stores, and application identities. The article names One API, New API, Veloera, Sub2API, and LiteLLM among gateway/proxy targets, and Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, and Ollama among AI application targets. This target list does not establish a common vulnerability or an affected release range.[0]
The liveness-bypass work reconstructed the verification client's face-detection events, position tracking, and movement coordinates so that generated facial movement could follow the requested challenge. Client-visible movement accuracy was insufficient: the reported server-side decision rejected every attempt despite completion of the sequence within seven pixels of the target.[0]
A separate session assembled a real-time deepfake pipeline using an open-source face-animation model on a single consumer-grade GPU and reached 11.5 frames per second. The article presents this as technical progress toward live-video impersonation, not evidence that either targeted exchange accepted the generated video.[0]
The forged-document work addressed image-consistency checks by constructing synthetic camera metadata whose dimensions matched the JPEG's actual dimensions and preserving original camera compression tables that ordinary editing and re-saving can replace. The report does not establish that these manipulated documents passed verification.[0]
Key takeaways
- ARTEX enabled a reinforcing credential-abuse workflow: stolen model access powered agents that searched for more access. The distinctive operational concern is automated repetition and persistence across familiar weaknesses.[0]
- Deepfake generation, movement accuracy, and realistic document metadata should not be conflated with successful fraud. The supplied account explicitly reports rejection of every attempt against the described vendor's server-side liveness decision.[0]
- Shared use of ARTEX does not establish a single named operator behind the campaigns or earlier incidents. Likewise, the reported 1,465 exposed consoles measure visible deployments, not confirmed malicious operators or compromised victims.[0]
- The evidence supplies no actionable attacker infrastructure, file hashes, specific vulnerable-version ranges, or patch identifiers. Product targeting and framework prevalence are useful context, but they are not substitutes for compromise indicators or version-specific vulnerability findings.[0]
- The intelligence picture is single-source at the supplied-evidence level. AWS, FBI, and Censys findings are reported through The Register rather than independently available source bodies, limiting corroboration and verification of the underlying observations.[0]
Defensive actions
- Restrict sensitive internet-facing services to trusted networks, private connectivity, or dedicated administrative access paths.: These are measures explicitly attributed to the FBI's alert. They address the exposed applications, cloud infrastructure, and partner or supplier systems highlighted in the warning.[0]
- Place enterprise applications behind zero-trust access policies and enforce the associated access permissions.: The FBI reportedly recommended this access model as part of reducing exposure to ARTEX-assisted attacks against externally accessible applications and services.[0]
- Eliminate default and hardcoded credentials from the exposed services and AI-related applications under review.: AWS specifically identified default or hardcoded credentials among the weaknesses through which the agents succeeded.[0]
- Remove exposed AI API keys and cloud credentials from code and public artifacts.: The observed campaign reused stolen keys to power further credential discovery, and AWS identified publicly exposed keys as an enabling weakness.[0]
- Require authentication on application interfaces and model-inference endpoints.: Unauthenticated interfaces and inference endpoints were explicitly identified as conditions that enabled agent success.[0]
- Address unpatched software in the exposed application and AI-service estate.: AWS included unpatched software among the successful access conditions. The supplied report does not identify particular CVEs, vulnerable releases, or fixed versions, so it does not support a product-specific patch prescription.[0]
- Retain authoritative server-side liveness decisions rather than treating completion of the visible movement challenge as sufficient verification.: In the reported attempts, precise movement completion did not defeat the vendor's server-side rejection. That distinction is the directly observed defensive success in the KYC campaign.[0]