3 out of 4 devs failed to catch dangerous AI-suggested commands

Name: 3 out of 4 devs failed to catch dangerous AI-suggested commands
Availability: InStock
Author: abdullaachilov

by abdullaachilov·Mar 26, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●SolidNiche GemSolve My Problem

Trap-based training catches devs who blindly approve AI commands before production burns.

Strengths

•Go proxy intercepts responses at wire level, zero changes to existing AI tool workflows.
•Trap commands are realistic terraform/shell examples that would actually cause outages.
•Open-source proxy builds trust—no one installs closed-source security tools from startups.

Weaknesses

•Quiz is lead gen for paid product, not the actual training platform itself.
•Security training category already has KnowBe4 and established enterprise players.

Post Description

Hey HN, I built this. AgentsAegis is a trap-based security training app. Think KnowBe4 for AI agents

The backstory: I'm a software engineer 14yoe, I use Claude Code daily. Sometimes I approve permission requests and only then read what I just approved Which is ironic as my primary spec is core back-end: security and work with big data. So I built this for myself to not become one of these stories: "Claude ran terraform destroy for production"

Concept is simple: Go proxy that sits between AI assistant and the API, intercepts responses, and occasionally swaps in realistic trap commands. If you approve blindly - you get caught. Sounds harsh, but again - I really dont want me or anyone to add 'caused 13 hours outage' in their resume. The proxy is obviously open source, i dont expect anyone to install something from closed-source repo of young startup: github.com/agentsaegis/go-proxy The quiz (link in the title) is the free version of that concept. Takes 2 minutes, no signup. Already has 80 takers - 75% scored C or D, average 6.5/10. Curious what HN thinks — both about the quiz itself and whether this is a real problem worth solving.