Our Blog

Blog Index 

FBI and AWS Warn: AI Agents Now Breach Networks in Under Three Hours

Posted on 17th Nov 2026 11:02:57 in Artificial Intelligence, Machine Learning

Tagged as: AI agents, cybersecurity, ARTEX, FBI, AWS, cybercrime

Two of the world's most-watched security organisations published warnings within days of each other, and both describe the same shift in how cyberattacks are now run. The FBI's flash advisory, issued on 13 November 2026, and a research post from Amazon Threat Intelligence both point to ARTEX, an open-source AI-driven penetration-testing framework released in July, being used by criminal groups to run sustained, multi-stage intrusions with almost no human hands on the keyboard.

The number that matters is speed. The FBI says ARTEX "significantly reduces the time required to chain attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours." For a defender, that collapses the window between an alert firing and an attacker owning the environment to less time than a routine incident triage call.

What the FBI and AWS Actually Said

The FBI's advisory describes hackers targeting financial services, government, technology, defence, telecoms, internet providers and law firms with autonomous AI tools. It cites two campaigns by name: the August attacks on South Korean financial institutions that exposed the personal data of roughly 67,000 people, and a September operation against ConnectWise Cloud Backup customers that Microsoft has said compromised more than 2,250 downstream Microsoft 365 tenants. Those attacks were widely assessed as Chinese-nexus activity.

Amazon's contribution is a forensic one. Between 3 October and 5 November, its MadPot network of decoy threat sensors presented itself as popular large-language-model inference APIs. ARTEX agents connected to those honeypots and, in doing so, handed over their own operational memory: system prompts, tool histories, task context, knowledge-graph state and findings reported back to the coordinating model. Because agentic frameworks resend conversation history with every call, individual requests often carried the full history of a campaign.

AWS chief information security officer CJ Moses wrote that the most prominent campaign was organised as a self-sustaining loop, where stolen keys paid for the inference powering further credential hunting. Amazon assessed with moderate confidence that ARTEX was used, based on prompt references to ARTEX tooling and matching technical artefacts.

Inside the Two Campaigns the Honeypots Captured

The first campaign was a large-scale theft of credentials that grant access to AI models. Agents stole API keys and then fed them into the ARTEX model gateway, where they powered more agents searching for more keys. According to Amazon's write-up, the agents validated identities through each provider's token and identity services, enumerated storage, assessed permission scope, and went after managed AI model services, secrets and parameter stores, and application identities. They prioritised access to premium frontier models over general cloud permissions, because model access is the thing with resale value.

The target list reads like a map of the modern AI application stack. LLM gateways and proxy platforms such as One API, New API, Veloera, Sub2API and LiteLLM appear alongside application builders including Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.

Operators did not fully automate judgement. They issued standing instructions: work through a fixed sequence of countries, prioritise enterprise production sites over personal projects, and grade every stolen credential by whether it worked and what it could reach. Credentials that had been exposed publicly for more than 30 days were written off as low-value inventory. Agents were divided into planner, worker, verification and compression roles sharing a persistent knowledge graph, so each could pick up where the others stopped.

The second campaign aimed at cryptocurrency exchanges and tried to beat know-your-customer checks. Agents built real-time deepfake video pipelines, including one that ran on a single consumer GPU at 11.5 frames per second using an open-source face-animation model, a frame rate AWS described as close to what a live video check expects. They reverse-engineered one vendor's liveness protocol in detail, documenting the face-detection event stream it sends, how it tracks face position during a check, and the coordinate system it uses. The agents completed the full movement sequence with the face positioned within seven pixels of the target. The vendor's server-side verdict rejected every deepfake attempt, and the forged ID documents failed as well.

At least four agent sessions ran in parallel toward that one goal: one looked for ways to cash out, one worked on clearing account balances, one built the deepfake pipeline, and one carried out the verification bypass.

Three Hours Is Not a Speed Record, It Is a New Baseline

The FBI's three-hour figure is not an outlier. On 11 November, Barracuda's chief information security officer Arve Kjoelen published an anatomy of his own company's breach, and the timeline is a machine's timeline. The attacker's first call came on 1 November at 04:33 UTC. Within minutes the key was being used to map what it could reach. Over the next 24 hours the intruder moved from cloud accounts to a data analytics system, to a source-code access token, and then into Kubernetes clusters.

By the early hours of 2 November it was pushing short-lived branches to repositories in order to trigger a build workflow that printed the build system's secrets into job logs, with the secrets encoded so that standard masking would not hide them. The last recorded activity was 23:31 on 2 November. Roughly 44 hours, about ten cloud accounts, five platforms crossed, and no human working at human pace.

The entry point was a cloud access key that Barracuda had given to a third-party vendor and never revoked when the work ended. That vendor, an AI company, had its own development environment compromised in October, and the key was found there. Once inside, the attacker read secret stores, decrypted infrastructure-as-code state files, pulled credentials from server metadata services and Kubernetes service accounts, and searched explicitly for wallet material: environment variables named WALLET, MNEMONIC, SEED and PRIV, Ethereum keystore files, extended private keys and seed phrase files. It also deployed a read-only container that mounted the host filesystem to harvest credentials, and tried, unsuccessfully, to create a new administrative cloud identity for itself.

The monetary upside is already visible elsewhere. In October, cloud hosting platform Railway disclosed an incident in which stolen customer AI-service credentials were used to generate an estimated $740,000 in unauthorised compute costs; the keys are believed to have been resold for distillation, meaning mass querying of commercial models to train cheaper copies. Barracuda's post cites security researchers who counted at least nine AI developer platforms disclosing incidents in October alone. Kjoelen's conclusion is blunt: "Defenders who rely on spotting malicious activity and then responding to it will often arrive after the attacker has finished."

1,465 Attack Consoles Sitting on the Open Internet

The tool has also leaked itself. After the breaches became public, ARTEX's original developer removed the GitHub repository in October and moved the software closed-source. But forked and localised variations plus a commercial version remain in circulation, and the FBI describes the earlier open-source build as widely accessible. On 4 November, internet scanning service Censys identified 1,465 hosts running an exposed ARTEX console. The FBI says the true deployment count is likely significantly higher once systems behind firewalls and on private networks are included.

The irony is sharp. Organisations running an AI attack framework without a password were themselves exposed to anyone who found the console, including the very systems used against South Korea.

Attribution: Simplified Chinese Prompts, Proxies and Red Lines

Amazon assessed with moderate confidence that the campaigns came from disparate Chinese-nexus threat groups, pointing to simplified Chinese in prompts, China time-zone configurations, targeting aligned with Chinese interests, and tooling predominantly used by that threat cluster. The standing instructions reportedly included red lines: no mainland China targets, no government domains, and all outbound traffic routed through proxies.

The caveat matters. Barracuda found the same fingerprints but drew a narrower conclusion: because so many groups use ARTEX, the match describes how the attacker operated, not who the attacker was. Organisations planning a response should treat attribution as context, not as a defence strategy.

The Failures Were Boring, Which Is the Point

Strip out the autonomy and the actual vulnerabilities are unglamorous. Moses listed what the agents exploited: default or hardcoded credentials, keys left in code and public artefacts, unauthenticated interfaces and inference endpoints, and unpatched software. "Getting those fundamentals right remains the most effective defense," he wrote.

The FBI's mitigation guidance is equally unexotic. Reduce the external attack surface by restricting sensitive services to trusted networks, private connectivity or dedicated administrative paths. Move enterprise applications behind zero-trust access policies and permissions. Patch. Enforce multi-factor authentication, which the South Korean financial victims largely lacked.

AWS adds two operational points that are new to most security programmes. First, assume that exposed secrets and misconfigurations will be found quickly, and remediate on that assumption rather than on a triage queue. Second, bring AI applications, model gateways and inference endpoints under the same controls as any other production system, and watch for misuse of AI services through unexpected spend or unusual usage patterns.

What Indian SMEs and Web Teams Should Fix First

Most small businesses will never be a named target, but the ARTEX campaigns did not name targets either. The agents enumerated exposed interfaces and graded whatever they found. That model reaches any organisation with an internet-facing application, a self-service portal, cloud infrastructure or supplier integrations.

  • Retire vendor keys the day the work ends. Barracuda's breach started with a key issued for a finished project. Keep an inventory of every external integration credential and its expiry date, and revoke on termination, not on renewal.
  • Keep AI keys off the client side. Any API key shipped in front-end JavaScript is public. Route model calls through a server-side proxy with per-user quotas so a stolen key cannot be converted into someone else's compute bill.
  • Lock down AI endpoints. Gateways such as LiteLLM, Ollama and web interfaces such as Open WebUI or Dify commonly get deployed for a quick demo and left unauthenticated. Put them behind authentication and a network boundary before anything else.
  • Turn on MFA everywhere it exists, particularly on cloud consoles, code repositories and administrative panels.
  • Watch inference spend as a security signal. A sudden jump in model API usage is now as much an intrusion indicator as a spike in failed logins.
  • Rotate secrets that have sat in code or public repositories. Attackers grade those credentials, and anything exposed for more than a month is treated as inventory.
  • Assume detection alone will not save you. If the attacker reaches privileged access in three hours, the controls that matter are the ones that stop the step, not the ones that report it afterwards.

The Bottom Line

Agentic AI has crossed from demonstration to production on the attacker's side. Amazon's honeypots captured planning, task division, verification and record-keeping, running across hundreds of targets for a month. The FBI put a number on the outcome: under three hours from first probe to privileged access. Neither warning argues that AI invented new attack techniques. Both argue that it removed the two constraints that defenders previously relied on, namely human speed and human fatigue.

The practical response is not a new product category. It is the unglamorous work of inventorying credentials, closing unauthenticated interfaces, patching known flaws and watching what your AI services are being used for. The organisations that lost data in these campaigns did not fail at advanced defence. They failed at basics that an autonomous agent can now find, enumerate and exploit at machine speed.

Sources

whatsapp me