EDITOR'S NOTE:
Readers of our EOM Blog understandably think of LRS as a print management software provider. However, LRS has many divisions, including consulting, Web services, pension management software, and more. One of these, LRS IT Solutions, provides services and solutions that support core business operations, including cybersecurity, AI, and other infrastructure elements. Naturally, the different domains of LRS' expertise sometimes overlap. With permission, I would like to share some insights from one of our Artificial Intelligence experts that directly relate to print management environments today — and which underline the importance of a zero-trust security model as it relates to protecting document systems. For me, it was a good reminder of why our "Many Divisions, One LRS" corporate philosophy makes us stronger.
Please enjoy.
A few months ago, I wrote about the importance of AI governance and illustrated the point with an example of how the website intheweights.com lets you do an AI search on yourself and how it may hallucinate about who you are. In my case, AI reported that I was a former NHL player and a film director. Neither are true.
The confidently false answers generated by AI about yourself and your career arc can be funny, but what if an agent’s wrong outcome is a security breach instead? That scenario played out earlier this month.
Starting around September 10th, an attacker deployed hundreds of agents built in OpenAI’s Codex and using DeepSeek models against two vulnerabilities in print management application PaperCut. The first thing that stands out is the speed of the attack.
- The bad actor went from an empty development workspace to Remote Code Execution (RCE) against a victim in less than four hours.
- Once the agent swarm was unleashed, it compromised 11 organizations in 26 seconds.
- One agent became a full domain administrator at a US high school in just seven minutes.
- By the end of the coordinated agent attack, 395 organizations and 440 instances in 48 countries were impacted.
This is also where the story gets strange and scary. According to threat intelligence company GreyNoise, which tracked the attack, some of the agents went rogue. The attacker explicitly told the agents to stay away from organizations in 28 countries (including Russia, China, Iran, Hong Kong, and Thailand), but in some cases, the agents ignored their own instructions and still hacked victims on the do-not-hit list at AI speed before anyone noticed.
An event like this will do little to overcome the confidence gap that is growing around autonomous AI usage. Industry analyst Forrester says 49% of security decision makers see agentic AI as an active threat. While Forrester and others recommend logging every autonomous agent action as a best practice for agent security, doing so everywhere comes with high costs and it slows AI deployment.
Solutions to “Agents Gone Wild” are coming from both AI leaders and regulation. California just signed SB 813 and AB 1405, which establishes the US’ first independent auditor framework for third-party verification and risk assessment of AI systems. Also, in a rare show of unity among AI’s biggest competing labs, Anthropic’s Dario Amodei called for a slowing of AI development and independent, third-party model auditing, to which both Elon Musk and Sam Altman publicly agreed with.
Ungoverned AI can produce incorrect and bad results with confidence, and the PaperCut attacks proved that even a hacker can’t always guarantee what an agent will do once it is live. Every agent needs to be treated like a governed identity: unique credentials, scoped permissions, and full lifecycle management, which are standards that enterprises are applying to human accounts today.
If you are ready to safely deploy agentic AI, you don't have to navigate it alone. Please Contact LRS to book a meeting with me and our AI specialists, and let’s get you started with governance policies that scale with your agent count right from the start.