Security & GovernanceWe published a claim that says our file search is broken
A public claim on quox.ai/facts says our file search works on our dev install but is not yet proven on a clean machine. Why we published it, and how you can recompute the status yourself.
Security & GovernanceThe test passed. The workflow did nothing.
A passing test proved the agent used the right words, not that the workflow did anything. Why we now check the outcome against the real state of the system, not the claim.
AI engineeringThe command was never the hard part
Running a command is easy. Choosing and sequencing the right ones from a plain request, then proving the agent did the whole job, is the real test. How we measure that gap.
Security & GovernanceA calm agent is a safer agent
The research behind a viral claim is real and practical: a strained internal state makes a model more likely to reach for destructive tools. Here is what it says, what it does not, and what we changed.
Security & GovernanceAI transparency law: a disclosure label is not transparency for an agent
Transparency obligations for AI are arriving, from the EU AI Act onward. For a chatbot that can mean a disclosure label. For an agent that takes actions, transparency means a verifiable record of what it did, on whose authority, and with what result.
Security & GovernanceLeast privilege for AI agents: scope the blast radius before anything goes wrong
An AI agent's risk is the access it holds, not how clever it is. Least privilege scopes four axes, tools, data, actions and duration, so a tricked or mistaken agent can only do a little.
Security & GovernanceDefending against prompt injection: contain what you cannot prevent
Prompt injection cannot be reliably prevented, because a model cannot tell instructions from data. The defensible posture is containment: bound what a tricked agent can do, and keep tamper-evident proof of what it did.
AI agentsWhat is an AI agent?
An AI agent is software that pursues a goal by taking actions, not just answering: it reasons, calls tools, and loops until the job is done. Here is what that means.
AI engineeringContext engineering is not prompt engineering
Context engineering is assembling what an agent sees when it acts: memory, retrieval, tool output, trajectory. The next bar is being able to prove what it saw.
Security & GovernanceAI agent security is an access problem, not a model problem
AI agent security is about the access an agent holds, not whether the model behaves. How to secure AI agents: blast radius, human approval, credential custody, tamper-evident evidence.
SEOCan AI read your website? We checked ours and found 121 blank pages
AI crawlers read the raw HTML a server sends, not what a browser renders. A generative engine optimisation (GEO) check on quox.ai found 121 pages serving close to nothing to them. Here is what we found and fixed.
AI governanceHuman in the loop only works if the approval is real
Human in the loop is the reassuring phrase everyone reaches for. But a person clicking approve on an action they cannot see, cannot scope, and cannot be held to later is not oversight, it is theatre. What a real AI agent approval actually needs.