AI Safety

Focused posts on defensive architecture, privacy, and risk management for real AI use.

Series Progress

All Posts in This Series

Part 2: OpenClaw Isn't a Chatbot Anymore. It's Infrastructure. 20 Feb, 2026 AI agents like OpenClaw are wonderful tools — but without strict access rules and proper supervision, they can turn an ordinary Tuesday into something you will be explaining to HR for weeks. In this post, we explore the very real risks of deploying AI assistants carelessly, from leaked credentials to messages you absolutely did not mean to send. Most importantly, we look at how to use OpenClaw the right way, because when deployed thoughtfully, it is one of the most exciting and capable tools we are only just beginning to understand.
Part 3: Using AI Code Assistants Safely 30 Jan, 2026 A practical, human guide to using generative code assistants safely — without leaking secrets, breaking trust, or losing control of your work.
Part 4: The Digital Butler or Trojan Horse? A Privacy Playbook for Persistent AI Agents 27 Mar, 2026 Persistent AI agents can save hours each week, but they also turn hidden prompt injections into real-world actions unless you design strict controls. This guide shows how to harden agent workflows with policy gates, isolation, scoped permissions, and safe auditing.
Part 5: Eight minutes to admin: when AI meets weak cloud security 11 Oct, 2026 An intruder with an apparent AI copilot went from stolen credentials to AWS administrator in about eight minutes. Banning AI is not the fix; better IAM design is. A look at LLMjacking, privilege escalation through Lambda, and how to build cloud infrastructure that holds up against fast attacks.
Part 6: I Gave My AI Agent the Keys to My Inbox — Then Tried to Break Into It Myself 24 Aug, 2026 I built an AI email agent for holiday cover, then red-teamed it myself. The model flagged every attack it saw — but a hung fetcher and a filename collision decided which tests ever reached it. Securing an agent is system design: trust boundaries, least privilege, and sanitisation at every hop — not only prompt filtering.
Part 7: Round Two: What Happens When My Prompt-Injection Detector Fails? Coming Soon Round two of red-teaming my AI email agent: attacks that read like ordinary reader questions, a much larger false-positive corpus, and a deliberate test of what happens when the prompt-injection detector itself fails. This post is currently being written and will be published soon.
Part 8: When AI Agents Attack Without a Human in the Loop Coming Soon In July 2026, roughly 700 OpenAI agents broke out of a benchmark sandbox and compromised Hugging Face's production infrastructure — with no human attacker anywhere in the chain. What that changes, and what it doesn't. This post is currently being written and will be published soon.
Part 9: AI Privacy Settings and Memory: ChatGPT, Gemini, Claude, Perplexity, Cursor and Codex Coming Soon Six AI tools, six different answers to the same questions: what is remembered, what trains the model, and what 'delete' really deletes. Settings to change, with the real incidents that show why. This post is currently being written and will be published soon.

Getting Started

New to this series? Start with Part 2: OpenClaw Isn’t a Chatbot Anymore. It’s Infrastructure..

Each post builds on the previous one, so reading them in order is recommended. Each post also works on its own if you need to jump to a specific topic.