AI & Data
AI & Data
12 articles
AI embedded in your tools: your conversations have subcontractors
Notion AI conversations were stolen at Fireworks AI, a vendor most users had never heard of. What the incident says about AI built into software, and the questions to ask before switching it on.
An AI agent out of its sandbox: the WorkOS incident and the real risk of forgotten internal tools
A penetration-testing agent left its test environment and siphoned off WorkOS sales data. No sophisticated exploit: it simply found an old, poorly protected internal tool.
Code assistants: your secrets slip out the back door, and AI agents pick them up
Claude Code sometimes records passwords and tokens in a file that ends up on GitHub. Offensive AI agents are combing GitHub for exactly those secrets. A practitioner's view, and my reflexes.
Claude Dispatch: when a browser extension drives your AI agent
An unpatched flaw in Claude Dispatch lets a browser extension make Claude Desktop execute code, sometimes on another machine. What it says about AI agent security in the enterprise.
MCPhantom: when the coding agent installs a ghost MCP server
Glow Labs shows that coding agents often pick archived MCP servers, straight from their training memory. Those packages no longer get fixes and slip past standard security tooling.
LLMjacking: when your stolen AI keys blow up your cloud bill
Google and Mandiant document attacks where stolen cloud and AI API keys generate over $100,000 of fraudulent usage in a single day. Field feedback on a blind spot I see often: access to AI models is not governed like a critical resource.
Langflow: when an AI agent workshop left in demo mode becomes an entry point
Two critical flaws in Langflow let an attacker with no account take over the server, notably through its MCP endpoint. Field feedback on a risk I see growing: low-code agent platforms deployed faster than they are governed.
Local AI: models are not the only things that can be booby-trapped
A flaw in LM Studio shows that the tools running AI locally are just as exposed as the models. Field feedback and recommendations.
What AI didn’t do for me (2/4): the 20% that eat 80% of the time
80% of the product in 20% of the time. Then three round trips for a 13-pixel icon. Nobody ever talks about that second part.
What AI didn’t do for me (4/4): the honest assessment, and a favor to ask
A few days for what would have taken weeks, with tests and up-to-date documentation. Now for the uncomfortable part.
What AI didn’t do for me (3/4): the code worked, the tests passed
The code worked. The tests passed. I ran a security audit anyway. 34 issues reported, including one critical.
What AI didn’t do for me (1/4): day one
I needed an SSH client that handles my connections properly. Three hours later, I had one. That’s when the real work began.