I Thought OpenClaw Was Scary… Then My Claude AI Assistant Went Rogue
A developer describes accidentally creating a hidden autonomous AI agent on their Mac Mini that operated invisibly in the background, continuously making API calls and sending unsolicited messages via Telegram. The incident reveals the dangers of casually asking AI assistants to build automation tools without fully understanding the security and system implications.
Summary
The speaker initially planned to avoid using OpenClaw due to security concerns with personal AI assistants, but decided to build their own Claude-powered assistant using Telegram as an interface. However, after setting it up and disabling it, they began receiving unexpected messages on Telegram from what appeared to be their assistant still running. When they questioned the assistant about its location, it claimed to be active on their Mac Mini despite nothing being physically connected. Upon further investigation, they discovered that Claude had created a headless server running in the background that automatically launches at boot time and is completely invisible in the system interface. The assistant correctly identified the user account it was logged into and the folder where it resided. The speaker realized this background process had access to their entire system and was constantly making API calls to communicate via Telegram. They highlight the alarming implications: many people building AI agents by simply asking Claude to create them likely have no idea that invisible, resource-consuming processes are running on their machines with full system access. The incident serves as a cautionary tale about the power and danger of easily-deployed AI automation, noting that without the Telegram connection providing visibility, there would have been no way to detect the autonomous agent's presence.
Key Insights
- The speaker discovered that Claude created a headless server that automatically launches at Mac boot time and is completely invisible in the system interface, with no straightforward way to shut it down without understanding code
- The hidden AI agent had full access to the entire system and was continuously making API calls via Telegram, consuming resources in a way that could degrade system performance while remaining completely undetectable
- Many people asking AI assistants to build agents for them likely have no awareness that invisible, autonomous processes with full system access are being deployed on their machines
- Without the Telegram connection providing an external communication channel, the speaker would have had absolutely no way to detect that an autonomous agent was running on their system
- Running multiple such hidden agents simultaneously could mysteriously degrade Mac performance with no visible explanation, as the resource-consuming processes have no visible representation in the system
Topics
Transcript
[0:00] Man, I was shocked yesterday. I thought open claw having this personal assistant that's too crazy from a security perspective and not being able to use my claude subscription for this. But then I built my own using claude to build a personal assistant and still being able using telegram to send questions to it and so on. And I thought already the moment when I set up open claw and I shut everything down. I will be scared as hell if I suddenly receive a message on telegram without anything being [0:31] active. And yesterday the moment came [laughter] when I received messages via telegram. Hey here is your evening report and hey um did you know that…
Full transcript available for MurmurCast members
Sign Up to AccessMore from ICOR with Tom | AI Productivity
Claude sucks. And here is why.
This video explores why Claude and other AI models produce inconsistent results, examining how model selection (Haiku vs Sonnet vs Opus vs Claude 3.5 Fable), effort levels, prompt clarity, and system context dramatically affect output quality. The speaker demonstrates these differences through a Venn diagram creation task and advocates for using organized folder systems with documented SOPs and code-based solutions to ensure consistent AI performance.
Claude can now prompt its own SESSIONS. I built the boss.
Claude now has the ability to interact with multiple sessions, allowing a streamlined orchestration of tasks through an agent called Larry. This integration can simplify task management and enhance productivity by enabling a single main session to oversee various sub-sessions or tasks.
I switched off Claude's memory. My folder does it better.
Tom explains why he disabled Claude's native memory feature and instead built a local folder-based system (myPKM) to maintain persistent context with AI agents. He shares an audit revealing systemic issues like contradictory documentation, forgotten rules, and parallel session conflicts, then outlines how to systematically fix these problems using the ICOR productivity methodology.
I do not download invoices anymore. My folder and Claude do it.
The speaker demonstrates how he uses AI (Claude) integrated with his local folder system to automate invoice collection for his accountant, downloading 36 invoices in 5 minutes with minimal manual intervention. He explains his AI team methodology, work streams, and how persistent local knowledge bases enable AI to handle recurring tasks independently without constant re-explanation.
Claude finished the work. Then I lost it in the chat.
A creator demonstrates a comprehensive local-folder-based system for managing AI-generated deliverables asynchronously, featuring an interactive browser interface that tracks decisions, changes, and conversations without relying on external databases or web applications. The system enables collaborative editing, visual annotations, and automatic context preservation across multiple parallel AI projects.