See what matters
in artificial intelligence.
Explore the research, tools, infrastructure, and decisions shaping AI. Read explainers and analysis with room for evidence, uncertainty, and a clearer view of what comes next.
Start with a practical guide
Follow a task from connecting tools to checking its result.
Latest perspectives
Browse all articles →-
Structured AI Output: Valid JSON Can Still Be Wrong
Separate JSON syntax, schema rules, factual evidence and permission checks before turning an AI-generated record into an action.
-
A Citation Is a Starting Point: Check an AI Answer Against Its Source
A practical source-checking worksheet for AI answers: verify the document, find the supporting passage, and match the claim’s scope to the evidence.
-
When a Document Gives an AI Agent Instructions: Understanding Prompt Injection
A practical explanation of indirect prompt injection, using a fictional document workflow to show where permissions, approval and output checks matter.
-
How to Evaluate an AI Agent: A Task Scorecard You Can Reuse
A worked example for checking an agent’s final result, permissions, recovery and consistency, with a reusable scorecard and clearly labelled hypothetical outcomes.
-
AI Reviewing Its Own Conversations: What Anthropic’s October Research Shows
A careful reading of Anthropic’s October 16 self-assessment research: the reported metrics, their limits, and questions for a monitoring rollout.
-
Gemini’s October 13 Update: What Changed and What to Check Before Switching
What Google announced on October 13, how to check your access, and a practical way to evaluate Gemini on reports, presentations and coding tasks.
-
PHP2Shell and AI-Assisted Security: What Site Owners Should Verify
AI-assisted discovery only helps site owners when fixes reach running systems. Here is how PHP2Shell connects research, official patches and hosting verification.
-
AI Cybersecurity: Model Supply-Chain Security in 2026
How model-loading formats and software versions affect supply-chain risk, and which provenance, isolation and permission checks to make before using third-party weights.
-
MCP and A2A Explained: Tool Access, Agent Communication and Trust Boundaries
MCP and A2A address different connections in an agent workflow. MCP connects an assistant to tools and resources; A2A supports communication between agents. Neither…
AI-assisted publishing. Clear responsibility.
The AI Prism uses AI for drafting, research assistance, and some illustrations. Content is credited to the publication, rather than fictional experts. Read our editorial standards or send a correction.