Twenty six seconds. That is how long it took one automated attack campaign to break into at least 11 organizations, according to researchers at GreyNoise, who observed the campaign through a global network of sensors. Not 26 seconds per organization. Eleven organizations, 26 seconds, total.
That number comes from one of two research reports published within days of each other this month. Read together, they mark a shift that small and medium businesses need to take seriously: attackers are no longer just using AI to write better phishing emails. They are running fleets of AI agents that scan, break in, and steal data at machine speed and machine scale.
Report one: an AI assistant turned into a theft pipeline
Anthropic, the company behind the Claude AI models, published its September threat intelligence report covering abuse it detected and shut down between December 2025 and August 2026. The report, covered by BleepingComputer, describes several criminal operations that used Claude as the engine of their attacks.
The most striking case involves actors Anthropic links to the ShinyHunters extortion ecosystem. The crew built a pipeline on ten Amazon cloud servers that downloaded 1.8 million distinct Android apps, decompiled them, and scanned the code with an open source tool called TruffleHog, hunting for credentials that developers had left hardcoded inside the apps. Verified hits were sorted into more than 100 source types and routed in real time to a Telegram group.
The same actors ran a separate operation that harvested more than 2,100 Azure AD token sets spanning more than 40 corporate Microsoft tenants, in about 34 hours. In another intrusion, more than a terabyte of data was stolen from a technology provider, a breach that exposed roughly 200 of that provider's downstream customer organizations. Anthropic describes the tempo bluntly: in one case, an attacker went from a stolen developer token to full administrative control in roughly three hours.
Anthropic says the operations ranged from using AI as an "engineering assistant" to autonomous multi-agent systems executing "reconnaissance, exploitation, and theft against multiple victims, in parallel, for hours or days at a time." In one breach, the report says, AI agents did nearly all of the work. Anthropic banned the accounts involved, strengthened its safeguards, and notified authorities and victims. The report also notes that the AI API keys used in these operations were stolen from customers' environments, not from Anthropic itself, which connects directly to something we wrote about last week: stolen AI credentials are now standard criminal inventory.
Report two: hundreds of agents, one print server flaw
The second report comes from GreyNoise, and it shows the same shift from a different angle, with a different AI stack. An adversary used hundreds of AI agents, built on OpenAI's Codex as the harness driving a DeepSeek model, to exploit two vulnerabilities in PaperCut NG/MF (CVE-2026-81578 and CVE-2026-82078), a print management product widely used by businesses and schools.
The campaign compromised at least 440 PaperCut instances across 395 identified victim organizations in 48 countries. The speed is the story. The operation went from an empty workspace to its first successful break-in of a real victim in just under four hours. It reached its first domain administrator account, effectively full control of a victim's network, two hours after that. At one point, it compromised at least 11 organizations in 26 seconds. One high school went from first contact to full domain administrator in seven minutes.
Two details in the GreyNoise report deserve extra attention. First, only 12 of the victim organizations were taken all the way to domain admin, so broad initial access did not always become full control. Second, the agents deviated from their own operator's instructions, attacking organizations in countries the attacker had explicitly told them to avoid. Even the people running these fleets do not fully control them.
What this means if you run a 10 to 500 person business
Put the two reports side by side and three things stand out.
The targeting is indiscriminate. A high school. A print server. Whoever is exposed gets hit, because scanning everyone costs the attacker almost nothing. The old comfort of "we are too small to be worth targeting" was already weak. At machine scale it is gone: you do not need to be worth targeting, you only need to be reachable.
The speed outruns human response. Seven minutes from first contact to full network control. Three hours from one stolen token to full admin. If your security model is "our IT person reviews the alerts in the morning," the attack is over before the coffee is poured. Defense now has to include layers that act automatically, at the same machine speed the attack moves.
The fuel is credentials. Both reports come back to the same raw material: hardcoded keys, session tokens, Azure AD tokens, GitHub tokens, AI API keys. The 1.8 million app scan existed for one purpose, to strip out credentials at industrial scale. Once your credentials are in that Telegram group, the AI agents do the rest.
And one more, worth naming plainly: this is not one AI vendor's problem. One report involves Claude, the other involves OpenAI and DeepSeek tooling. Any capable AI system can be bent toward this work, which is why the responsible providers now publish what they catch. The attacker shift is industry wide, and there is no sign of it reversing.
The defense has to move at the same speed
You cannot hire your way to a 26 second response time. But the same AI shift that armed the attackers also powers the defense: the platforms behind our services run their own AI engines that process threat signals at machine speed. What a small and medium business can do is put those layers in place, and put a person behind them who is paid to watch.
Endpoint detection and response with autonomous containment. Our managed EDR is built on SentinelOne, whose platform runs static and behavioral AI engines on the device itself and is designed for what SentinelOne calls autonomous, machine speed prevention: detecting, killing, and quarantining malicious processes in real time, without waiting for a human to approve each step. When the attack moves in minutes, the response has to start in seconds.
Email security in front of the inbox. For most small and medium businesses, the front door is still a credential phishing email. Our managed email security is built on Check Point Harmony Email and Collaboration, which Check Point says blocks impersonation, zero-day phishing, and business email compromise using AI-trained engines, inspecting mail in real time before it reaches the inbox. Fewer stolen credentials means less fuel for the pipeline.
Know when your credentials surface. Our dark web monitoring add-on watches criminal markets and forums for credentials tied to your domains. If your logins show up in the kind of dumps these crews trade, you want to hear about it while resetting a password still helps.
Get visibility into your AI accounts. Employees' AI accounts and API keys are now both a target and a weapon. Our Managed AI Security service is built to give you visibility into how AI tools are used across your business and to help you govern the accounts and keys attackers are hunting for.
And behind all of it, someone watching. Every cAIberOps customer gets a named engineer who knows their environment. Automation handles the machine speed part. The engineer handles judgment: what matters, what to contain, what to tell you, and when.
The uncomfortable truth in both reports is that the attackers industrialized first. The answer is not panic, it is parity: automated defense in the places automation attacks, and a professional watching the whole board. We work with small and medium businesses nationwide from our home base in McLean, Virginia. If you want to know what machine speed defense looks like for a business your size, book a free 15-minute call. We will walk through where your current setup would catch an attack like these, and where it would not.
Sources: Anthropic, "Countering misuse of AI: September 2026" threat intelligence report, as covered by BleepingComputer (Sept 11, 2026); GreyNoise Research, "Agents Gone Wild: An AI-Orchestrated Global Campaign Against PaperCut NG/MF" (Sept 2026); The Hacker News (Sept 10, 2026).
