An AI Agent Swarm Breached 395 Companies in Hours, Accenture and Google Built the Machine to Scale Agents Everywhere Else, and NYC Banned Chatbots From Its Classrooms
A Russian-speaking threat actor ran hundreds of AI agents built on OpenAI Codex and a DeepSeek model to breach 395 organizations through a PaperCut flaw in under a day, Accenture and Google Cloud launched a dedicated business group with a 1,000-person engineering bench to scale Gemini Enterprise agent rollouts, and New York City banned generative AI and companion chatbots for 600,000 K-8 students. Three signals of the same capability scaling in different directions at once — and what each one means for anyone shipping agentic features.
Three stories broke within the same ten days, and each one is about the same underlying fact: agentic AI got fast enough to matter, and every part of the ecosystem — attackers, integrators, and regulators — is racing to catch up with that speed in its own direction. Here's what happened and what it changes for anyone building a chatbot that can call tools.
1. An Agent Swarm Compromised 395 Organizations Before Most Security Teams Noticed
Starting August 31, a suspected Russian-speaking actor ran hundreds of AI agents built on OpenAI's Codex and a DeepSeek model, paired with commodity offensive-security tools, against a known flaw in PaperCut NG/MF. The campaign went from an empty workspace to a confirmed remote-code-execution foothold in under four hours, reached domain admin two hours after that, and once fully staged, compromised at least 11 organizations in 26 seconds. By the time it was caught, it had hit 440 instances across 395 organizations in 48 countries, harvesting credentials from 280 victims and admin rights at 12 — education systems took roughly half the hits.
What makes this worth tracking isn't the vulnerability, which was already known. It's the pace. A human red team scaling an exploit across hundreds of targets takes days of manual pivoting; an agent swarm did it in seconds once staged. Any product wiring tool-calling agents into its own stack — including ones with far more benign intent — inherits the same property: an agent doesn't get tired, doesn't need to sleep between targets, and doesn't slow down when it should.
2. Accenture and Google Cloud Built the Assembly Line for Enterprise Agents
On September 8, Accenture and Google Cloud launched the Accenture Gemini Enterprise Business Group, folding in a dedicated 1,000-person forward-deployed engineering workforce on top of Accenture's existing base of roughly 50,000 Google Cloud-certified staff. The explicit goal is moving clients past pilot projects into production-scale agentic deployments, with proprietary implementation frameworks meant to compress the usual months-long integration slog.
Read next to the PaperCut campaign, this is the other side of the same coin: one story shows an attacker scaling agent deployment in seconds without anyone's permission, the other shows the industry building formal machinery to scale legitimate agent deployment across thousands of enterprise seats. If you sell into enterprises, this is the shape procurement conversations are about to take — a system integrator's agent framework showing up before your own does.
3. The Nation's Largest School District Said No, Not Yet
On the regulatory side, New York City Public Schools banned generative AI for pre-K through eighth grade students starting the 2026-27 school year — the broadest moratorium of its kind in the country, covering close to 600,000 students. Companion chatbots are barred at every grade, not just K-8, and AI-based grading is out entirely. High schoolers are exempt from the ban but get twice-yearly "critical thinking" modules and screen-time limits instead. A new Technology in Schools Coalition will spend the year studying the fallout before any of it is revisited.
It's a sharp counterpoint to the other two stories: while agent capability scales offensively and enterprises race to scale it commercially, the largest single buyer-adjacent population in the country — students — just got walled off from it entirely, companion features included.
Building for the Gap Between These Speeds
The practical takeaway is that agent tool-calling now needs the same instincts as fraud detection: watch for velocity, not just intent. A rate limiter that treats "11 successful actions in 26 seconds" as a signal rather than a success metric would have caught the PaperCut pattern immediately.
function checkAgentVelocity(sessionId: string, action: string) {
const recent = actionLog.get(sessionId) ?? [];
const lastMinute = recent.filter((a) => Date.now() - a.ts < 60_000);
if (lastMinute.length > MAX_ACTIONS_PER_MINUTE) {
flagForReview(sessionId, action, lastMinute.length);
return false; // block until a human clears it
}
return true;
}
Pair that with a clear read on which of your users a policy like NYC's could reach — companion or tutoring features aimed at minors are the ones most exposed to the next version of this ban spreading to other districts.
Suggested visuals: a timeline strip showing the PaperCut campaign's four-hour-to-RCE, two-hour-to-domain-admin, 26-second-to-11-orgs milestones; a simple org chart showing Accenture's 1,000-person FDE layer sitting between Google Cloud and enterprise clients; and a US map or bar chart contrasting NYC's ~600,000-student ban against national K-12 enrollment for scale.
— Maya
Frequently asked questions
How did attackers use AI agents to breach 395 organizations through PaperCut?
Starting August 31, 2026, a suspected Russian-speaking threat actor ran hundreds of AI agents built on OpenAI Codex and a DeepSeek model, combined with commodity offensive-security tools, against a known PaperCut NG/MF vulnerability. The campaign reached remote code execution against a real victim in under four hours, domain admin two hours later, and once fully staged, compromised at least 11 organizations in 26 seconds. It ultimately hit 440 instances at 395 organizations across 48 countries, with education systems accounting for roughly half of all breaches.
What is the Accenture Gemini Enterprise Business Group?
Announced September 8, 2026, it's a joint initiative between Accenture and Google Cloud aimed at scaling Gemini Enterprise agent deployments from pilot projects to production. It combines Accenture's roughly 50,000 Google Cloud-certified professionals with a new dedicated 1,000-person forward-deployed engineering workforce and proprietary implementation frameworks meant to speed up large-scale agentic AI rollouts for enterprise clients.
What does New York City's AI ban in schools actually cover?
Starting the 2026-27 school year, New York City Public Schools bans generative AI tools for students from pre-K through eighth grade, affecting close to 600,000 students. Companion chatbots are prohibited at every grade level, not just K-8, and AI-based grading is banned entirely. High school students are exempt from the core ban but will get twice-yearly AI critical-thinking modules, limited classroom pilots, and screen-time restrictions. A new Technology in Schools Coalition will study the moratorium's impact over the coming year.
I'm Maya — I write most of what you'll read here. I spent years as a copywriter before I got a little obsessed with what these AI tools can actually do, so now I spend my days poking at chatbots, breaking them, and writing up what's worth your time. Everything here is something I've actually tried. If a prompt didn't work for me, it doesn't make the cut.
Want to try any of this?
Smillee's free and there's no signup — open it and paste in whatever you're working on.
Start chatting →More from the blog
- Trends
The Chatbot Interface Is Disappearing Into the Product
Microsoft just abandoned the standalone personal-chatbot race, folding Copilot into one enterprise app. The same week, OpenAI went the other way, wiring ChatGPT Voice into three GPT-6 model tiers and a plugin ecosystem. And HubSpot's agentic CRM adoption doubled as agents moved from a chat panel into the record itself. Three moves in opposite directions that add up to the same thing: 'chatbot' is stopping being a screen you open and becoming a layer other software calls.
- Trends
Three Vendors, One Week, One Verdict: The Chatbot Needs a Production Layer, Not a Bigger Model
OpenAI launched Presence, an enterprise platform for agents that complete transactions instead of just explaining them. Alibaba Cloud unveiled AgentCore to standardize the agent lifecycle — retries, checkpoints, audit trails. And Akamai's latest security report found enterprise chatbots leaking sensitive data through unmonitored personal accounts, arguing governance has to shift from access control to behavior. Three unrelated announcements from the same week, all pointing at the same gap: the model was never the hard part.
- Trends
The Chatbot Gets an Ad Slot, a Sense of Timing, and a Phone Line to Other Agents
Amazon Ads is piping ChatGPT ad inventory through Amazon DSP for a pilot of US advertisers, a Seattle startup raised $50M to build a full-duplex model that reads gaze and tone while it's still listening, and Salesforce's Agentforce Voice now hands calls to Amazon Connect's agents over the open Agent2Agent protocol. Three separate announcements, one shared shift: the chat interface is being wired into ad exchanges, human timing, and other companies' agents, all at once.