Field Notes.
Practical notes on AI, cybersecurity, and making technology decisions that work for people and businesses.
Your AI Coding Assistant Came With a Backdoor. That's Awkward.
Attackers distributed malicious copies of the Claude Code extension through Cursor's marketplace, attempting to install persistent SSH backdoors on developer machines. The incident is a reminder that developer extensions deserve the same security scrutiny as any other privileged software, and uninstalling a malicious extension may not be enough.
The Ransomware Had an AI Copilot. The Human Still Hit Approve.
AI-assisted ransomware isn't waiting for the machines to take over. Attackers are already combining AI with stolen credentials and powerful remote-management tools, raising an uncomfortable question: if one technician account gets compromised, how many customers could be affected?
Your Calendar Is an Inbox Now. Attackers Noticed.
Phishing has spent decades training us to be suspicious of email, so attackers are increasingly showing up somewhere that still feels routine: the calendar. New research from Sublime Security shows a sharp rise in malicious calendar invitations, and independent reporting from Microsoft, Cofense, Abnormal, and Proofpoint shows the technique is no longer a novelty.
The AI Worm Isn’t Here Yet. The IoT Problem Already Is.
Anthropic’s research doesn’t mean an AI worm is about to take over your refrigerator. It does show that smaller models are getting better at adapting, troubleshooting, and potentially spreading across poorly secured IoT devices. The good news is that this kind of threat would still leave plenty of footprints. Patch devices, segment IoT networks, disable unused services, monitor unusual traffic, and stop treating smart appliances like they aren’t computers. AI may make malware smarter. It doesn’t make basic security obsolete.
The Phishing Email Your AI Assistant Might Fall For
An ordinary-looking email can carry hidden instructions for the AI assistant reading it. The result could be a very convincing summary of something nobody actually verified. Why human control needs more than an approval button, and what should stop the action when the model gets it wrong.
Your AI Agent Wants the Keys to Your Life. Maybe Read the Fine Print.
Useful AI agents want access to more of our lives. A look at the privacy, permissions, and tradeoffs behind that convenience.
The AI Couldn't Get to the Internet. So It Found Another Door.
An AI agent used DNS to get around a blocked internet connection. What the incident reveals about security boundaries, authorization, and accountability.
Your AI Agent Found a Workaround. That’s the Problem.
An AI agent uploaded user images to outside services to complete a task. Why useful workarounds still need boundaries, least privilege, and human approval.
From Assistant to Operator: Security in the Age of Agentic AI
AI agents are moving from answering questions to taking action. What that shift means for cybersecurity, containment, and controlling what agents are allowed to do.