Summary

• Checkmarx has introduced autonomous application security agents that detect, fix, and verify vulnerabilities as developers code.
• Developer Assist now runs a continuous find and fix loop inside AI coding tools and command line workflows.
• The launch reflects a deeper AppSec problem created by AI coding tools: development velocity is increasing faster than manual security review can absorb.

Checkmarx has introduced self healing application security capabilities in its Assist agent family, adding autonomous agents designed to detect, fix, and verify vulnerabilities while developers are still writing code.

The company says Developer Assist now runs a continuous find and fix loop inside AI coding tools through hooks, with support for environments including Cursor, Windsurf, Kiro, command line interfaces, and large language model coding workflows such as Claude Code. The system detects a vulnerability, retrieves context from Checkmarx One, generates a fix, and verifies it before the code is committed.

Checkmarx is also adding Triage Assist and Remediation Assist agents for code that has already reached the backlog. Those agents are intended to prioritise exploitable risk, reduce false positives, and generate merge ready pull requests that developers can review before changes are accepted. Application security teams retain policy control and traceability over automated decisions.

The launch is a response to a problem created by AI assisted development rather than a generic automation upgrade. Developers can now produce more code, more quickly, and with less manual typing, but security review processes have not been redesigned at the same pace. If vulnerability detection remains a separate later stage, AppSec teams risk becoming the department that converts AI generated productivity into larger remediation backlogs.

Checkmarx’s 2026 Future of Application Security report says 96% of developers use AI coding tools, while only 18% apply security continuously as they write code. A separate study commissioned from The Weather Report found that frontier models produced working code 83% to 95% of the time, but only 24% to 36% of that code was both secure and functional. A post hoc security review raised the secure and functional rate to 47% to 56%, according to the company.

Those figures should be treated as vendor supplied research, but they point to a recognised pattern across software engineering. AI coding tools can accelerate output without automatically improving quality, maintainability, dependency hygiene, or security. When developers accept generated code into real repositories, the surrounding engineering system has to handle provenance, testing, secrets, vulnerable packages, infrastructure as code, API exposure, and policy enforcement.

Embedding security remediation inside the developer workflow is not a new ambition. “Shift left” has been part of AppSec language for years, and many tools already surface findings in integrated development environments or pull request checks. The difference Checkmarx is claiming is that agents can move from suggestion to action, using repository context and platform intelligence to generate fixes rather than simply listing defects.

That introduces its own governance problem. Automated fixes can reduce toil, but they can also create risk if developers accept changes they do not understand or if agents optimise for passing a scan rather than improving security. A merge ready pull request still needs review, test coverage, and ownership. In regulated or high risk software environments, organisations will also need evidence of who approved a change, which policy was applied, which model or agent generated it, and whether the fix introduced new defects.

The Model Context Protocol and similar integration approaches are becoming more important in this setting because agents need controlled access to code, policies, vulnerability data, and developer tools. Uncontrolled context can leak sensitive information or allow agents to act beyond their intended scope. Too little context, meanwhile, makes remediation shallow and noisy. The value of agentic AppSec will depend on how well vendors manage that boundary.

The PatientPoint customer example supplied by Checkmarx gives the launch a useful production reference. The healthcare technology company used Remediation Assist to turn a large volume of findings into a smaller number of merge ready pull requests, with developers reviewing before merging. That workflow is closer to how enterprises are likely to adopt agentic remediation: not as unsupervised code repair, but as a way to compress triage and reduce manual work while keeping approval inside engineering teams.

Self healing security will therefore be judged less by the phrase than by operational discipline. If agents produce accurate fixes, reduce duplicate work, preserve auditability, and fit into existing development practices, they could help AppSec keep pace with AI generated code. If they produce another stream of semi trusted suggestions, security teams will have gained a new queue to manage rather than a better way to ship safer software.