AI Code and Prompt Injection Expand the Attack Surface
Security teams are now facing a combined input-output problem. AI-written code increases the volume of code that needs review, and indirect prompt injection turns ordinary web content into a way to poison AI agents that read it. The standard response of trusting the model or adding a quick review step misses that both sides of the pipeline can be manipulated.
ProjectDiscovery says only 38% of cybersecurity practitioners feel they are keeping up with the AI-driven code load, and nearly 60% say review is getting harder. Google says indirect prompt injection is a top priority and that it swept the public web for known patterns, because AI systems can silently follow malicious instructions embedded in websites, email, or documents. The report also shows defenders are most worried about corporate secrets, unreliable dependencies, and business logic flaws in AI-generated code.
The risk is not just more bugs. It is more untrusted input reaching both code and agents at machine speed, which raises the odds of poisoned instructions and exploitable application logic landing in the same environment.