The distilled version
- A hacker used a swarm of AI agents, mostly running on their own, to break into over 100 websites including those of John Deere, SharkNinja and Cummins, per security firm Norforge and Reuters.
- Separately, Microsoft warned that AI agents deployed inside companies can become an ‘agentic insider’ risk if they are given too much access to data, tools and systems.
- Both stories point to the same lesson: any AI agent or third-party tool connected to your systems needs limited, checked permissions, not blanket trust.
What happened
Security firm Norforge found that a hacker broke into more than 100 websites using AI agents that operated with little human involvement, according to The Verge and Reuters. Reuters reported the victims included John Deere, SharkNinja and Cummins, with the attack built on an open-source AI platform called Cairn, where users set a start and end goal and the agents work out the steps. Norforge found more than 20 exposed Cairn instances it considers likely malicious, with the most active attacker’s successful breaches each taking about 23 hours, targeting companies in the US, Canada, Australia and Japan.
John Deere told Reuters it found no indication sensitive information was accessed. Cummins said it is investigating a potential cyberattack on a website run by a third-party vendor.
On the same day, Microsoft published a blog post arguing that AI agents companies deploy internally can become what it calls ‘agentic insiders’: agents that are often overprivileged, have access to data and systems, and can act unpredictably even without being hacked. Microsoft says organizations will need what it calls Guardian Agents, an established industry term, to watch and moderate every action of every agent, and points to its Project Perception system as a step toward that.
Why it matters
Small businesses increasingly connect AI agents and tools to email, CRMs, e-commerce platforms and support software. Both stories show the same gap: agents are often given broad access and then left to run with little oversight, which is exactly what attackers and Microsoft both describe as the weak point.
You don’t need to run your own AI agents to be affected. If a vendor or contractor connects an agent to systems that touch your data, their setup becomes your risk too.
What it means for your business
If you run the business: ask any vendor or staff member who has connected an AI agent or automation to your systems what permissions it has, and whether those permissions are limited to what the task actually needs.
What to do
Act now- Make a list this week of every AI agent, chatbot or automation connected to your business systems, including ones set up by vendors.
- For each one, check what data and systems it can access, and remove any permission it doesn’t need for its actual job.
- Ask vendors directly whether their AI agents run with restricted, monitored access or broad standing access.
- Add a review step: someone checks what an agent did, at least weekly, rather than assuming it’s working as intended.
Act now: worth doing this month.
Jargon in this story
Sources
- 1A hacker broke into John Deere and SharkNinja with an army of AI agents (opens in a new tab) · The Verge · News report · Oct 19, 2026
- 2Securing the agentic insider (opens in a new tab) · Microsoft · Company announcement · Oct 19, 2026