What happened
Amazon Threat Intelligence says its MadPot network of decoy sensors captured requests from AI agents between October 3 and November 5, 2026. The requests included agents’ operating memory from offensive campaigns run on ARTEX, an open-source agentic AI penetration-testing framework. MadPot presented interfaces resembling popular LLM inference APIs; ARTEX agents sent prompts to them as if they were backends. The records included persistent task context, prompts, tool histories, knowledge-graph state and reported findings.
AWS assesses with moderate confidence that ARTEX was involved, based on prompt references and technical artifacts. It also assesses with moderate confidence that the activity involved disparate Chinese-nexus groups, citing simplified Chinese prompts, China time-zone settings and targeting patterns. Amazon says it notified potentially affected AWS customers where appropriate. It could not confirm activity against other organizations, and shares indicators of compromise for defenders to review their own environments.
The largest campaign described sought credentials for AI models. Agents searched internet-exposed applications, extracted upstream provider keys, tested them and stored working credentials in a managed pool. Stolen keys were then fed into the framework’s model gateway to power further credential hunting. The operators prioritized frontier-model access and treated cloud credentials as a path to additional permissions and services. Targets included LLM gateways such as LiteLLM, One API, New API, Veloera and Sub2API, along with AI application builders. The playbook mapped products to attack paths including uninitialized installations, anonymous management endpoints, default administrator credentials and known vulnerabilities. AWS specifically names LiteLLM’s default master key among the credentials involved.
The other campaign targeted know-your-customer checks at two cryptocurrency exchanges. Agent records describe attempts to create real-time deepfake video for live selfie checks and forged identity-document images. One video attempt followed a liveness check’s face-position sequence to within seven pixels of its target, but the vendor’s server-side verdict rejected every recorded attempt. The agents also described a video pipeline reaching 11.5 frames per second on a consumer-grade GPU. At least four sessions worked concurrently on different parts of the operation and shared findings through cross-session tools.
AWS says the techniques were not new; the change is the speed and persistence agents can bring to them. Its defensive recommendations are to treat AI applications, model gateways and inference endpoints like other production systems, remove default or hardcoded credentials, and monitor for unexpected inference spend or unusual usage. The post includes indicators of compromise with observed dates spanning October 14 to November 2, 2026.
What the archive adds
New Runtime reported that Censys found 1,465 publicly exposed ARTEX AI penetration-testing consoles in its earlier coverage on November 7, 2026. This AWS finding is a new fact on that known story: it describes agents’ reported campaigns, rather than exposed consoles alone. New Runtime also covered a self-service Bedrock-plus-LiteLLM gateway in an earlier note on November 13, 2026; the new security finding is directly relevant to operating that kind of gateway.
Trend signal
This does not yet fit a pattern New Runtime tracks. AWS’s account supports a practical security warning about gateway exposure, but it does not establish a broader shift in where agent security boundaries are moving.
Why now
The observation window ended on November 5, and AWS published its findings on November 13, 2026. The timing matters because the post gives operators indicators and concrete remediation guidance while describing campaigns that AWS says were sustained and financially motivated.
Connections to other events
New Runtime’s November 7 coverage said Censys had found 1,465 publicly exposed ARTEX consoles. AWS’s account adds reported offensive activity conducted through ARTEX agents, while cautioning that MadPot observed the agents’ records rather than the underlying campaigns as they happened. Together, the accounts connect exposed agent tooling with the operational activity attackers may pursue through it, without proving that the exposed consoles were the same systems used in these campaigns.
Why it matters
A model gateway concentrates access: one interface may hold credentials for several downstream providers. If its management surface is exposed or protected by defaults, attackers can potentially turn a single weakness into model usage paid for by someone else, and use that access to continue searching for more credentials. That makes gateway configuration an operational security issue, not just an integration detail.
For teams, the useful response is broader than checking whether LiteLLM is installed. Inventory internet-facing AI applications and inference endpoints, verify that management interfaces require authentication, remove default and hardcoded secrets, and rotate credentials that may have been exposed. Treat gateway usage and spend as production telemetry. AWS’s account also shows why agent-enabled operations can persist across sessions and coordinate work, but it does not prove that every reported action succeeded. Defenders should use the indicators and claims as leads to check their own environments, not as proof of compromise.
Limits and uncertainty
AWS says it did not observe the campaigns’ activity on the systems described in agent records; MadPot captured requests sent to decoy inference interfaces. The post therefore cannot independently verify the agents’ reported outcomes or confirm the full scope of activity against other organizations. Attribution to ARTEX and to Chinese-nexus groups is moderate-confidence, not definitive. AWS does not identify all affected organizations or establish that the KYC bypass attempts succeeded; it says the recorded deepfake attempts were rejected. The observation period and published indicators are not a complete measure of current exposure.
Skill to build now
This week, inventory every internet-facing AI gateway and inference endpoint your team operates. For each one, confirm that administrative and provider-configuration routes require authentication; remove default credentials and hardcoded keys; rotate any secret that may have been exposed; and review usage for unexpected inference spend. If an endpoint must be public, document its owner and the controls protecting its management plane.
Reader takeaway
The central lesson is not that agents discovered a new exploit. AWS says familiar weaknesses—exposed interfaces, weak credentials and leaked keys—supported persistent campaigns that could recycle stolen AI access. Secure gateways as production systems, rotate exposed secrets, and monitor both access and inference costs.