A security-focused analysis finds that AI-generated code frequently ships with vulnerabilities, with Veracode's Spring GenAI Code Security Update showing 45% of outputs containing a known vulnerability across Java, JavaScript, C#, and Python, while security pass rates linger around 55% for two years.

To address this, CAST (Code-Aware Security Transformation) proposes a security gateway that enforces policy inline as the model writes code, routing editor and IDE traffic through a local gateway to inspect both outbound tokens and inbound responses in real time.

CAST targets four surface areas of failure beyond the code itself: prompt injection, unsafe tool calls, data leakage on outbound paths, and actions taken without trace, with the scan focusing on the token stream rather than the finished outputs.

The gateway is not a replacement for traditional security measures like SAST or human review, but aims to reduce risk during the critical generation window when code is still being produced, arguing that post-generation review can be too late for certain vulnerabilities.

Cencurity is open source under Apache-2.0, with its code on GitHub and a Product Hunt launch, signaling a practical, auditable, in-editor security control approach for AI-assisted development.

Vulnerability distribution varies by weakness class, with insecure crypto and SQL injection showing high failure rates (86% and 82%), while log injection is much lower (13%), indicating training data and pedagogy shape model security behavior.

Key design decisions for Cencurity include: never exposing API keys outside the IDE; logging only policy violations to avoid creating new breach surfaces; enforcing security inline to prevent vulnerable code from completion; operating locally via loopback to minimize latency and exposure; and supporting multiple providers/endpoints such as OpenAI, Anthropic, and Gemini.

Stanford studies reveal a paradox: users with AI assistants write less secure code and grow more confident in its security, suggesting tool-assisted code can feel safer even when it's not, though skeptical prompts and rewriting can reduce vulnerabilities at the cost of scalability.