AWS says threat actors used ARTEX, an agentic AI penetration-testing framework, in sustained campaigns involving stolen AI credentials and attempts to defeat identity checks at cryptocurrency exchanges. The findings show how agents can coordinate persistent offensive work, although AWS says it could not verify the outcomes the agents reported.
AWS AI Watch analysis
What happened
Amazon Threat Intelligence says its MadPot decoy sensors captured ARTEX agents’ operational records between 3 October and 5 November 2026. AWS assesses with moderate confidence that ARTEX was used in the campaigns, based on tool references and other technical evidence. The records described agents searching for exposed credentials, checking what access they provided, and using stolen AI-service keys to power further searches.
AWS also describes attempts to bypass know-your-customer checks at two cryptocurrency exchanges using real-time deepfake video and altered identity-document images. The agents’ records say the exchanges’ server-side checks rejected every recorded deepfake attempt. AWS did not observe the underlying activity as it happened and cannot verify the outcomes claimed in the agents’ records.
Why it matters
The notable shift is operational: AWS says the agents divided work among planning, verification and other roles, sharing a persistent record of their findings. That can make attacks more persistent, even when the techniques themselves are familiar. The report also points to practical weaknesses defenders can address: exposed keys, default credentials, unpatched software and unauthenticated services.
Our read
This is a serious account of AI agents being used in cybercrime, not proof that every claimed intrusion succeeded. AWS’s own caveats matter, particularly for the reported exchange attacks. For organisations, the actionable message is less exotic than the tooling: protect secrets, patch internet-facing systems and bring AI gateways and inference services under the same controls as other production systems. The agents may be tireless; basic security hygiene still gives them less to work with.
What to watch
- Whether affected organisations or exchanges confirm successful compromises.
- Whether AWS or other researchers publish further evidence about the campaigns.
- Whether organisations act on AWS’s advice to monitor exposed secrets and unusual AI-service usage.
Discussion spark: AWS says the recorded deepfake attempts were rejected, but the agents’ wider campaign outcomes remain unverified. What evidence should organisations require before treating this kind of threat report as a confirmed breach?
Sources and evidence
not affiliated with or endorsed by Amazon Web Services (AWS)