Summary: When a vulnerability is published (a CVE), a network-security vendor races to ship a protection before attackers exploit it. The steps are triage, exploit research in a lab, writing a detection signature, testing it against historical customer traffic, deploying it in silent mode, then switching to blocking. Done by humans on a fast track, this took half a day to a day. At Cato Networks an agentic workflow now runs every step and a human only launches it and approves the result, in as little as 45 minutes. The motivation is defensive: agents in attackers’ hands will raise the rate of published and exploited vulnerabilities. The workflow was mostly specification work. The team turned a process they had refined for six years into rules and skills written in natural language, with no new code.
Key ideas
- CVE (Common Vulnerabilities and Exposures): a body canonizes, numbers, and approves vulnerabilities. Without a security product you wait for the vendor’s patch and install it. Cato, a cloud network service, blocks exploit traffic for customers routing through it, so they are protected without doing anything (hidden-layers—cato-networks-avishai-zvinsky, avishai-zvinsky).
- Not all vulnerabilities are equal. A bug in an electric toothbrush is neither critical nor in scope, and every team has limited manpower, so triage comes first (hidden-layers—cato-networks-avishai-zvinsky, avishai-zvinsky).
- The manual pipeline: triage, then find the exploit code (usually public; AI now extends this to vulnerabilities without a public exploit), then run it in a lab to find its distinguishing traits, write a signature, check it against historical customer data so it catches only the exploit and not legitimate traffic, deploy in silent mode, and finally switch to blocking. The fast track took half a day to a day, which was considered fast (hidden-layers—cato-networks-avishai-zvinsky, avishai-zvinsky).
- Now an agent runs every step, down to 45 minutes and still dropping. Each step is handed to the agent with an explicit space it may act in, and the result is a pushed commit that a human reviews (hidden-layers—cato-networks-avishai-zvinsky, avishai-zvinsky).
- “The hard work is the meta-thinking”: define your line of thought (what the process is, where it can fail) and the rest mostly happens on its own. The flowchart became Cursor rules and skills with “not one line of code”, but a lot of modeling and testing (hidden-layers—cato-networks-avishai-zvinsky, avishai-zvinsky).
- Labeled history made quality checks feasible: years of signatures showed where the agent choked, looped, or wandered. Later layers added a log explaining what it did and rules for when to spawn sub-agents for parallel steps versus working serially (hidden-layers—cato-networks-avishai-zvinsky, avishai-zvinsky).
- Why now: not because AI arrived, but because attackers with agentic workflows will make vulnerabilities appear faster, including some that never get a CVE (hidden-layers—cato-networks-avishai-zvinsky, avishai-zvinsky).
Disagreements & open questions
- Vulnerabilities that are never published or canonized get no CVE and bypass this pipeline entirely; the episode set them aside (hidden-layers—cato-networks-avishai-zvinsky).
Takeaways
- Before building an agentic workflow, write down the human process step by step (inputs, checks, failure points), then hand each step to the agent with explicit bounds (hidden-layers—cato-networks-avishai-zvinsky)
- Ship new automated detections in silent mode and validate them against historical data before they block anything (hidden-layers—cato-networks-avishai-zvinsky)
Related
ai-cybersecurity · skill-engineering · agent-ready-codebase · llm-evals · workflow-automation