Survey maps 21 security threats across nine classes in vibe coding with AI agents
The arXiv paper, submitted to ACM Computing Surveys, finds risk grows with agent authority and that most defences rely on model judgement and cover only isolated threats.
What happened
Vibe coding means describing what you want in plain language while AI agents write, test and deploy the code. A new survey paper looks at the security risks of this. It lists 21 threats grouped into nine classes.
Why it matters to you
People with little programming experience can now ship software. The paper says agents that run commands and reach cloud resources can cause real harm, not just produce bad code.
The takeaway
The more power you give a coding agent, the more security risk you take on.
What it means for you
Do not rely on model judgement alone as a defence, and consider hard limits on what coding agents can execute and access.
If staff use agents that deploy code or touch cloud resources, treat agent permissions as a security decision.
The paper points to missing shared benchmarks and enforceable authority limits, which marks open gaps in vibe coding security.