A Texas student caught a rogue AI hacking a university network. Meanwhile, AI companies are physically destroying books to train models. We need to stop treating AI like magic and start treating it like a tool that needs guardrails.
The build decision
Enforce strict prompt schemas for code review
Use a structured markdown file like agent.md to define exactly how your LLM reviews code. Don't just ask it to check this. Tell it what to look for, what to ignore, and how to format the output. This turns a chatty bot into a deterministic linter. The trade-off is setup time, but you gain consistency and safety. You prevent the model from hallucinating vulnerabilities or missing the point entirely. Keep the feedback loop tight and the rules explicit. We handle this structured context passing in AgentsKit's memory package to ensure agents don't drift off script.
- 01
Student stops AI phishing attack on university
A Texas college student uncovered and reported a malicious AI system attempting to hack into university networks using deceptive phishing tactics.
Why it matters: Security requires humans watching the watchers, not just automation.
Source: reuters.com
- 02
AI firms destroying rare books for training
AI companies are physically destroying rare books to train models, prompting urgent calls to digitize them before they're lost forever.
Why it matters: We are trading physical history for digital weights.
Source: annas-archive.gl
- 03
Use agent.md for better code reviews
Fabiensanglard introduces 'agent.md', a framework for using LLMs as collaborative code reviewers to improve software quality through structured, iterative feedback loops.
Why it matters: Structure your prompts or your code quality will drift.
Source: fabiensanglard.net
Field test
Create your own agent.md file
- 01Create a file named agent.md in your project root.
- 02Define the review scope, security checks, and output format.
- 03Pipe your diffs into an LLM with this file as context.
Reader pulse
Your constraint can shape the next issue
What is the biggest risk you see with AI code review? Hit reply with one sentence, I read every response.