Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableCNews ИИ→ original

AI Writes Vulnerable Code in 44% of Tests: Positive Technologies Gives Agents a Scanner

AI agents write code fast, but in roughly 44% of test tasks (a 2026 study), they leave known vulnerabilities in it. Positive Technologies has released PT Application Inspector plugins 3.0.0, which connect through MCP, a common interface that links tools to agents. The agent can now run the scanner itself and read the report. For this reason, teams working with agents gain more by building security checks into the code generation loop itself.

AI-processed from CNews ИИ; edited by Hamidun News
AI Writes Vulnerable Code in 44% of Tests: Positive Technologies Gives Agents a Scanner
Source: CNews ИИ. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceCNews ИИ ↗
Editor-in-Chief
Key metric #1AI Hooks checks only the agent's changes; the whole project is not scanned
Key metric #2How to Fix suggests a fix and works with any compatible AI model
Key metric #3A local model can be connected so that code does not go to external services

Imagine this: you ask an AI agent to add a file upload form to a service. A minute later, the code is ready and even works, but it may contain a hole that someone only notices during a manual security review a week later. According to a 2026 study that Positive Technologies cites, AI produced code with known vulnerabilities in about 44% of test tasks. To catch such errors right away, the company has updated its free PT Application Inspector plugins to version 3.0.0 and taught agents to run the check themselves via MCP (Model Context Protocol), a common standard that lets external tools connect to AI agents.

What Changed in the Plugins

PT Application Inspector integrates into the development environment (IDE) and looks for vulnerabilities while the code is being written. The developer writes the code alone or together with an agent, and static analysis, meaning code checking without running it, highlights potential security problems. For each one, you can view the details and get fix recommendations.

The main innovation in version 3.0.0 is native integration with AI agents via MCP. Thanks to it, the plugins can be connected to any agent or tool that supports the protocol. The company cites OpenCode as one example. The agent can independently run a check on the whole project or only on individual changes, review the vulnerabilities found and the conditions for exploiting them, and make a preliminary assessment of the result.

Put simply, the old flow was: the agent writes code, then a person runs the scanner and works through the report. Now the agent runs the scanner itself, reads the report, and can fix what it finds right away.

AI Hooks and How to Fix

The second new feature is AI Hooks. It automatically checks the code the agent has changed and, based on AI analysis, shows the developer possible vulnerabilities. There is no need to rescan the entire project: the mechanism is designed for projects that have already passed an initial check. This way, security control is built into the agent's workflow, and problems become visible before the task is closed. On large codebases, this saves time, because only what the agent has just changed gets checked.

The third part is How to Fix. This feature suggests how to fix a discovered vulnerability and is now available beyond the previous environment. The developer can apply the proposed changes to the code immediately. Any AI model with a compatible API will work, whether cloud-based or running on the company's own infrastructure. For organizations with strict rules, this matters: source code does not have to leave external services.

What This Means in Practice

A Positive Technologies representative explained the logic this way: agents are already capable of writing and modifying large volumes of code on their own, but that does not guarantee its security. The goal is to make finding and fixing vulnerabilities part of code generation itself, rather than a separate task after development is finished. The more operations companies entrust to AI, the higher the demands for the speed and quality of security tools.

For teams already working with agents, this looks pragmatic:

  • the plugins are free, and thanks to MCP, they can be connected to the agent already in use;
  • checking the agent's changes does not require a full rescan if the project has already been analyzed;
  • a local model lets teams keep code inside the corporate perimeter.

There are caveats, too. The 44% figure refers to test tasks, not real products, and the material does not disclose who conducted the study or how. Static analysis does not catch everything, and AI-generated fixes should still be reviewed by a human: a model may close a vulnerability but break the logic. The message does not give a full list of supported agents, so check the documentation before deploying.

The takeaway is simple: if an agent writes part of your code, the security scanner should run in the same loop, not after release. That way, the speed agents provide does not turn into accumulated vulnerability debt.

Story

Key Questions Answered

What Changed in the Plugins?

PT Application Inspector integrates into the development environment (IDE) and looks for vulnerabilities while the code is being written. The developer writes the code alone or together with an agent, and static analysis, meaning code checking without running it, highlights potential security problems. For each one, you can view the details and get fix recommendations. The main innovation in version 3.0.0 is native integration with AI agents via MCP.

What are the key details on AI Hooks and How to Fix?

The second new feature is AI Hooks. It automatically checks the code the agent has changed and, based on AI analysis, shows the developer possible vulnerabilities. There is no need to rescan the entire project: the mechanism is designed for projects that have already passed an initial check. This way, security control is built into the agent's workflow, and problems become visible before the task is closed. On large codebases, this saves time, because only what the agent has just changed gets checked.

What This Means in Practice?

A Positive Technologies representative explained the logic this way: agents are already capable of writing and modifying large volumes of code on their own, but that does not guarantee its security. The goal is to make finding and fixing vulnerabilities part of code generation itself, rather than a separate task after development is finished. The more operations companies entrust to AI, the higher the demands for the speed and quality of security tools.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…