Key points
- AI models introduce several times more vulnerabilities per line than human developers.
- Malicious open-source packages account for most zero-days, growing over 600% in 2026.
- Snyk data shows eight new flaws introduced for every one fixed as of October 2026.
AI-driven software development is generating code faster than security teams can review it, creating a significant vulnerability backlog. According to Snyk, current coding models introduce several times more flaws per line of code than human developers. Many of these defects involve business logic, access control, and authentication, areas where traditional pattern-based scanners have limited visibility. This imbalance means that by the time security reviews occur, the risk profile of the application has already shifted dramatically.
In plain English
The core issue is that automated coding agents do not assess the safety of the dependencies they install. A malicious package can steal credentials during the installation phase, which happens before a pull request or continuous integration scan runs. Because these agents operate continuously, they can introduce harmful components without any human oversight. This allows attackers to embed theft mechanisms directly into the development environment before traditional security controls even activate.
The background
Most security teams were already managing a latent risk backlog before AI development accelerated. Snyk’s data from October 2026 showed roughly eight new vulnerabilities introduced for every one remediated. Attackers are now using AI to identify weaknesses and combine lower-severity issues into more serious exploits. Furthermore, malicious packages now account for the majority of open-source zero-days, with more than 20 high-profile attacks recorded in 2026 alone.
What changes now
Snyk argues that pull request scanning alone is no longer sufficient to address every risk. Application security must run in three places: checks inside the coding workflow, fast deterministic scans on every change in CI, and deeper analysis of the full application at regular intervals. This approach requires independent validation of proposed fixes rather than relying solely on the model that generated them. Teams must define clear policies governing what gets blocked, fixed, or requires human review.
The attacker's path
Teams usually face exposure when automated agents accept unvetted dependencies. The weakness relies on the gap between code generation and security verification. Attackers exploit this by injecting malicious logic that executes during installation, bypassing later-stage scans. This technique serves as an initial access vector, allowing credential theft before the code reaches human review or CI pipelines. Check whether your agents validate package integrity before installation.
What to do and how to stay safe: Snyk
- Verify that coding agents are restricted from installing dependencies without prior security assessment.
- Implement checks within the coding workflow to block unverified packages before they reach CI.
- Ensure proposed fixes undergo independent validation rather than accepting model-generated changes blindly.
- Monitor for credential theft indicators during the installation phase of new dependencies.
Step-by-step guide: Emergency Patching: Correcting Five Common Operational Errors
General security guidance from the RedTeam Log newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How many new vulnerabilities are introduced compared to those fixed?
Snyk data from October 2026 shows roughly eight new vulnerabilities introduced for every one remediated.
What type of flaws do AI models introduce more frequently?
Current models introduce several times more vulnerabilities per line than human developers, often in business logic and access control.
How do malicious packages bypass traditional security scans?
They can steal credentials during installation, which occurs before a pull request or CI scan runs.
