FBI Warns AI Agent Framework ARTEX Used in Cyber Intrusions, Including Crypto Wallet Searches

Daily Feed
FBI Warns AI Agent Framework ARTEX Used in Cyber Intrusions, Including Crypto Wallet Searches

FBI warns AI-agent framework ARTEX was used in cyber intrusions

The FBI says attackers used ARTEX, a framework built for authorized security testing, in intrusions across multiple sectors. Separate reports from Microsoft, Google, Barracuda and Amazon describe credential theft, cloud access and attempts to target cryptocurrency-related systems. But the reports do not show that every attempt succeeded or that the incidents involved the same operator.

  • ARTEX agents can handle multi-step tasks based on ordinary-language instructions.
  • Microsoft reported attempted sign-ins to more than 2, 250 Microsoft 365 tenants, not confirmed breaches.
  • A cloud key that remained active after vendor work ended enabled access to Barracuda’s environment.
  • Attackers searched for cryptocurrency wallet files. Barracuda said none were on its compromised servers.
  • Amazon found deepfake identity-check attempts in records sent to its decoy systems, but their full reach is unclear.

What ARTEX does and what it does not prove

ARTEX was built for authorized penetration testing. An operator can give its AI agents instructions in ordinary language, and the agents can carry out sequences of tasks that might otherwise require someone to handle each step.

That can speed up work such as reconnaissance, or gathering information about a target, and searching for credentials. It does not make the agents infallible, prove an attempted intrusion succeeded or reveal who was operating them.

In a November advisory, the FBI said ARTEX-related attacks had affected financial services, government, defense, healthcare and other sectors since at least August. The bureau also cited the theft of personal information belonging to approximately 67, 000 people in South Korea. Available information does not specify what data was taken or explain how that incident relates to the other reported campaigns.

Different reports, different evidence

Microsoft’s October investigation attributed one campaign to Storm-0940, which it described as a China-state-aligned group. Microsoft said the group used ARTEX against hundreds of organizations.

Microsoft also reported attempted sign-ins to more than 2, 250 Microsoft 365 tenants using stolen application credentials, login details assigned to software or services. The attempts followed the ConnectWise compromise. The number refers to targeted tenants, not confirmed account takeovers, and available information does not establish how the ConnectWise compromise relates to the Storm-0940 campaign.

Microsoft described follow-on activity in multiple customer environments, including reconnaissance and changes to application credentials, user identities and privileged roles. Privileged roles have elevated permissions that can give users administrative control over accounts or systems.

In a separate password-spraying operation, Microsoft recorded more than 151, 000 failed logins against 327 accounts across eight organizations. It observed no successful authentication. Password spraying means trying a limited set of passwords against many accounts. Microsoft’s findings do not establish that this operation was part of the same activity as the tenant sign-in attempts.

Google’s October research described ARTEX activity affecting organizations in the United States and Europe, including semiconductor, telecommunications and enterprise software businesses. Google observed financially motivated attackers and suspected espionage groups using the framework, but did not establish that all the activity came from one group.

Google also reported operational errors. In one intrusion, agents generated 240 log events, and investigators saw signs that separate agents interfered with each other’s work. The details show that agent-driven activity can leave a substantial record and get messy. On their own, they do not show how easily defenders could detect or stop it.

A cloud key left active, and a search for wallet files

Barracuda said an intrusion began with a cloud access key it had given to an AI vendor. The vendor’s development environment was compromised in October, and the attacker apparently obtained the key. Barracuda had not revoked it after the vendor finished its work. The key was used to access Barracuda’s cloud environment on November 1.

Over roughly 44 hours, the intruder reached around 10 cloud accounts, following credentials from one system to another. Barracuda said the sequence of activity, identifying traces and built-in credential-hunting routines matched ARTEX. It also cautioned that identifying the tool did not identify the person or group behind the intrusion. Similar access failures have featured in dark-web cybercrime investigations.

“That is on us.”

Barracuda used those words to acknowledge that it had not retired the vendor’s access key. The lapse is a reminder of a basic security risk: an attacker with a valid credential may not need to get past a more sophisticated defense.

The intruder searched compromised servers for Ethereum keystore files, extended private keys and seed phrase files. A keystore file can protect access to an Ethereum account. Extended private keys and seed phrases can provide access to wallet keys or help recover a wallet. Barracuda said none of that wallet material was on the compromised servers. The search shows what the attacker wanted, not that any cryptocurrency was stolen.

Deepfake attempts and crypto-exchange identity checks

Amazon’s November findings described attempts to bypass identity verification at two cryptocurrency exchanges using deepfake video and altered identity documents. Amazon based its assessment on agent records sent to its decoy systems between October 3 and November 5. It assessed ARTEX involvement with moderate confidence and said it could not confirm activity against every organization named in those records.

In one recorded verification flow, a verification vendor’s server rejected every deepfake attempt, including attempts in which an agent completed the required face-movement sequence. Amazon’s findings do not establish whether the vendor’s attempts were tied to either exchange or whether every attempt against the exchanges failed.

The result is encouraging, but limited: at least one server-side check rejected the recorded attempts. That does not show that every exchange or identity system can withstand deepfakes.

Amazon also described a separate campaign that collected credentials for AI services and fed stolen keys back into the attackers’ infrastructure to pay for more AI processing. The finding points to one way compromised accounts could help fund further automated activity.

The familiar weakness behind the new tools

Across the reports, the problems that enabled intrusions included exposed credentials, unpatched software and interfaces that did not require authentication. AI can speed up the work, but basic access controls still matter. A cloud key that remains active after a vendor’s work ends is a door left open.

GitBook separately told customers in November to replace integration secrets after an AI-agent attack accessed infrastructure credentials. Potentially exposed information included API tokens, Git Sync secrets and questions submitted to GitBook Assistant. Available information does not establish that the GitBook incident was connected to the ARTEX campaigns described by the other companies.

Key questions and answers

  • Were more than 2, 250 Microsoft 365 tenants breached?

    No. Microsoft reported attempted sign-ins to more than 2, 250 tenants using stolen application credentials. That number is not a count of confirmed breaches.

  • Did attackers find wallet material at Barracuda?

    No. Barracuda said the attacker searched for wallet-related files, but none was on the compromised servers.

  • Did the deepfake attempts beat identity checks?

    One verification vendor’s server rejected every recorded attempt Amazon described. The findings do not confirm the outcome of every attempt against the two unnamed exchanges.

  • Does ARTEX identify the attacker?

    No. Microsoft attributed one campaign to Storm-0940, but Barracuda said matching activity to ARTEX did not identify who carried out its intrusion. The reports do not show that all incidents involved the same operators.

  • What should organizations do?

    Revoke vendor credentials when work ends, protect application and cloud keys, patch exposed systems and require authentication for interfaces. Automation can speed up an intrusion. Removing stale access closes one of the easiest routes in.

Share this article

Powered by ADBYTES

Advertise smarter.

Adbytes.Media is a transparent advertising network where advertisers reach real audiences and publishers, affiliates & everyday members earn ADBYTES tokens. Join the community and start earning today.

Back to Blog