The keys, floating in the dark.
By Callahan Blackwood | November 16, 2026
The attack is not coming from a human operator hunched over a keyboard at three in the morning. It is coming from software. It runs at machine speed, it does not tire, and it was built specifically to find the credentials that unlock artificial intelligence infrastructure. This week, two separate disclosures confirmed that the weapons are real, the targets are known, and the question everyone in enterprise security should be answering is about accountability, not technical details.
API Keys Are the New Crown Jewels, and the Agents Know It
Amazon has uncovered an active campaign in which AI agents were deployed to target LiteLLM, Dify, and similar applications with one specific goal: stealing API keys. The Hacker News reported the finding on November 16, 2026. LiteLLM and Dify are not obscure research tools. They are production-grade middleware that sits between an organization's applications and the AI models those applications call. They route requests. They manage authentication. They hold the credentials that determine who gets to use which model, at what cost, and at what scale.
That makes them exactly the kind of target a well-designed attacker goes after. An API key extracted from a router-level application does not give you one account. It may give you access to every model the organization uses, every token budget it has negotiated, and every workflow that depends on those models running. The attackers running this campaign understood that. Amazon's finding is that autonomous agents were deployed to systematically hunt credential stores in AI infrastructure, not that someone tried a phishing email.
LiteLLM and Dify sit in the class of tools that grew quickly because developers needed them. They offered abstraction, routing, and a single interface over multiple AI providers. Organizations adopted them fast, often faster than their security teams could audit the deployment. The credential exposure that creates is structural, and this campaign found it.
Why This Campaign Design Matters Beyond the Headline
The significance goes beyond one campaign targeting one class of tool. It lives in what the campaign design reveals. Using AI agents to attack AI infrastructure is a closed loop. Agents searching for API keys that control agents. The credential theft compounds because the stolen access is itself agent-capable. An attacker with a harvested key from a production LiteLLM instance may have programmatic access to language models with enough context and permission scope to continue the intrusion further, automatically.
No specific victim organization was named in the Amazon disclosure reported by The Hacker News. That is not unusual in responsible disclosure contexts, and this article does not speculate beyond what the verified record confirms. The record confirms the method, the targets, and the class of access being sought.
The FBI Issued a Warning. That Is Not a Routine Event.
On November 15, 2026, BleepingComputer reported that the FBI issued a warning about a tool called ARTEX. The warning covers global cyberattacks that ARTEX is powering. The FBI does not issue global cyberattack warnings about every new piece of malware that surfaces. When the bureau issues one, it means the tool has already caused documented harm at scale, attribution is solid enough to name the tool publicly, and the threat is continuing.
ARTEX is an AI hacking tool. It is a purpose-built offensive capability that uses AI to automate the attack chain, not a kit assembled by amateurs adapting open-source scripts. The FBI's warning makes it a matter of federal record that this tool is active, that it is targeting organizations globally, and that it represents a threat serious enough for the bureau to alert the public.
That warning landed one day before Amazon's disclosure about API key-hunting agents. Two independent threads converged in the same 48-hour window. The timing is coincidental. The pattern is not.
How ARTEX Actually Works Inside a Network
GBHackers News reported on November 16, 2026 that ARTEX AI agents use Nmap, PsExec, and WinRM to automate enterprise network attacks. Each of those three tools deserves a sentence.
Nmap is a network scanner. Security teams use it to map what devices and services are running on a network. Attackers use it for the same reason: reconnaissance. ARTEX deploys it to understand the target environment before moving further.
PsExec is a remote execution utility. It allows commands to be run on remote systems across a Windows network. Administrators use it for legitimate maintenance. Attackers use it to move laterally, pushing code or commands from one compromised machine to another without touching the filesystem in ways that trigger alerts.
WinRM is Windows Remote Management. It is Microsoft's built-in protocol for remote administration. When an attacker gains the right credentials, WinRM gives them an authenticated shell on a remote Windows machine. No exploit required. Just valid credentials and an open protocol.
The pattern here is not about zero-days. ARTEX does not appear to rely on undiscovered vulnerabilities in these tools. It relies on the fact that these three capabilities, combined and automated, can traverse an enterprise network with minimal human supervision on the attacking side. The agent scans, finds, executes, and moves. It does this across multiple targets, in parallel, without the attacker needing to issue each command manually. That is the operational advantage AI brings to offense: scale without proportional labor.
This matters for incident response as well. Human attackers make timing mistakes, take breaks, generate patterns. An agent running Nmap, then PsExec, then WinRM in an automated sequence does not. The behavioral signature is different. Organizations relying on anomaly detection tuned for human attack patterns may miss the tempo entirely.
Defence Is Responding, But Catching Up Is Not the Same as Being Ahead
Horizon3, a cybersecurity firm, launched a dedicated lab for AI cyber defence research, IT Brief Australia reported on November 16, 2026. The timing is not incidental. Building a dedicated research lab for AI-driven defence in the same week that an FBI warning and an Amazon disclosure both confirm AI-driven offense is industry responding to an operational reality, not a theoretical one.
A lab launch is not a deployed solution. It is the beginning of organized work. Horizon3 is signalling that the firm believes AI-native defence will require purpose-built research infrastructure, not just incremental updates to existing security tooling. That is the right framing. Existing endpoint detection, network monitoring, and SIEM configurations were built around human attacker behavior. An AI agent running Nmap-to-PsExec-to-WinRM chains at machine speed, across dozens of targets simultaneously, is a different problem class.
What the lab will produce, on what timeline, and whether it will close the gap with tools like ARTEX: none of that is established by the available record. What is established is that a named firm committed resources to building AI-specific defensive capability at the moment it is most needed.
The UK Is Spending Money. Some Are Not Convinced.
New AI cyber funding in the United Kingdom was welcomed this week, but resilience doubts linger. IT Brief UK reported that assessment on November 16, 2026. The specific fund amount and the government body disbursing it are not established by the information available in the verified pack for this article, and this piece does not reach beyond it. The available record does establish the nature of the reception: the money was not rejected, but practitioners did not treat it as sufficient.
That reaction pattern is familiar after high-profile threat disclosures. Funding arrives, often with the implicit message that the government takes the threat seriously. The people who actually run the defences look at the operational picture and see a gap between the size of the problem and the size of the investment. The doubts about resilience that IT Brief UK describes this week fit that pattern exactly.
The UK question is part of a wider policy arc. AI is a dual-use technology at a civilizational scale. The same capabilities that make AI agents useful for routing API calls, automating workflows, and running network diagnostics also make them useful for attacking those same systems. Governments that want to benefit from AI productivity need to invest in the defensive capability that matches the offensive capability adversaries are already deploying. Funding that arrives slowly, in amounts that prompt resilience doubts, is funding that is behind the curve.
The Accountability Vacuum at the Center of All of This
This week's disclosures force a question: when an AI agent steals credentials, or runs reconnaissance, or executes lateral movement across a network, who is accountable for the harm it causes?
The human who built the tool is one answer. The organization that deployed it is another. Neither of those answers maps cleanly onto the actual harm chain in the cases reported this week. The Amazon disclosure involves agents deployed autonomously to hunt credentials. The FBI warning involves a tool powering global attacks that has already caused documented harm. By the time attribution happens, if it happens, the agent has run, the keys are gone, and the access has been used.
Accountability in cybersecurity has always lagged the threat. Legislation takes years. Prosecution requires attribution. Attribution requires forensics that often come too late to matter for the victim. AI agents compress the attack timeline in ways that make the accountability lag worse. The agent can complete the credential theft, the lateral movement, and the data exfiltration in a window so short that the organization does not know it happened until hours or days after the fact.
The tools in ARTEX's toolkit illustrate the point. Nmap, PsExec, and WinRM all appear in the logs of well-run security operations centers as normal traffic. An AI agent using them in sequence to attack is doing something that looks, in each individual step, like routine network administration. The malicious pattern only becomes visible when you examine the sequence, the timing, and the target. That analysis takes time. The attack does not.
Who Was Watching the Keys
The API key story is worth pressing specifically. LiteLLM and Dify are not secret tools. They are widely documented, openly deployed, and used by organizations that often adopted them under time pressure, with engineering teams focused on making AI workflows run rather than on auditing credential exposure.
Amazon's finding that agents were deployed to target these specific applications means that whoever designed the campaign had already mapped the architecture. They knew where the keys live. They knew which tools hold them. They designed agents specifically to extract them. That is not opportunistic. That is targeted, automated theft of a specific class of credential that grants access to AI infrastructure at scale.
The organizations running LiteLLM and Dify did not necessarily do anything obviously wrong. They used tools that worked. The problem is that the tools that work are now the targets. The question of who watches the keys is not rhetorical. It lands on specific roles: the security team that approved the deployment, the engineering lead who chose the tool, the CISO who signed the architecture review. When the attacker is software and the credential store is an AI routing layer, those humans are still accountable for the architecture that created the exposure. That accountability does not dissolve because the attack was automated.
The FBI warning about ARTEX puts a federal name on the threat. Amazon's disclosure puts a named class of targets on the attack surface. Horizon3 is building the lab. The UK is spending the money, however cautiously the practitioners are receiving it. But the gap in all of this is structural governance: the policies, audit requirements, and assigned ownership that determine who is responsible for credential security in AI infrastructure before the agent arrives, not after.
That gap does not close with a lab launch or a funding announcement. It closes when organizations treat API keys to AI routing middleware with the same rigor they apply to privileged account credentials in traditional IT. It closes faster when regulators, insurers, and legal frameworks catch up to the reality that the attacker is now software, running at scale, using tools that look like normal traffic until they do not.
The agents are the attack now. Catching up starts with admitting that the accountability model built for human attackers does not fit the problem anymore.
Sources
- Amazon uncovers AI agent campaign targeting LiteLLM, Dify for API keys – The Hacker News, November 16, 2026
- FBI warns ARTEX AI hacking tool powers global cyberattacks – BleepingComputer, November 15, 2026
- ARTEX AI agents use Nmap, PsExec, WinRM to automate enterprise attacks – GBHackers News, November 16, 2026
- Horizon3 launches AI cyber defence research lab – IT Brief Australia, November 16, 2026
- New UK cyber AI funding welcomed but resilience doubts linger – IT Brief UK, November 16, 2026

