Why does AI make a mature AppSec program more important?

AI demands mature application security because it increases how much code ships, lowers how well each line is understood, and adds new risks such as prompt injection and over-privileged agents. If your review, testing and response processes were barely coping before, more code arriving faster will break them.

What changes when AI writes code?

Volume goes up. A developer with an assistant opens more pull requests, and each is larger. Reviewers have the same hours. Findings from scanners rise with code volume, so triage queues grow.

Understanding goes down per line, since the author did not type it. That makes review and testing the real quality gates.

Then there is new surface. Applications that call language models take untrusted text as input to something that follows instructions. OWASP's Top 10 for LLM Applications lists prompt injection, sensitive information disclosure and excessive agency among the risks, and these are application security problems, not only model problems.

How do you measure maturity?

OWASP SAMM, the Software Assurance Maturity Model, organises security work into five business functions: Governance, Design, Implementation, Verification and Operations. Each has practices scored on maturity levels, so you can see where you are thin.

Run a short self-assessment. For each function ask one blunt question.

• Governance: Is there a written policy for AI tool use and who owns it?
• Design: Do new features get threat modelled, including AI components?
• Implementation: Are dependencies controlled and secrets kept out of code?
• Verification: Do tests and scans run on every change, with agreed blocking rules?
• Operations: Can you find, fix and redeploy within days of a disclosure?

Where does AI stress each function?

A constructed example for a mid-sized product team that adopts an assistant in January.

By March, pull requests per week have doubled. Governance has no policy, so engineers use three different tools. Design reviews are skipped because "the assistant already thought about it." Verification catches fewer problems per review minute because diffs are bigger. A scanner backlog of 900 findings grows, and a real exposure sits under noise. In Operations, nobody can say which services include a library named in a new advisory.

None of these gaps is caused by the assistant. It exposed existing ones faster.

What should you fix first?

Verification and operations pay back quickest.

1. Make scans and tests run on every change and fail on clear conditions.
2. Use reachability to cut the backlog to what code actually calls.
3. Keep an inventory and SBOM per release so you can answer disclosure questions in an hour.
4. Write the AI use policy, including which data may be shared and which models are approved.
5. Add AI components to threat models, covering untrusted input, tool permissions and output handling.

Safeguard provides dependency and SBOM scanning, reachability analysis, policy gates, MCP Guardrail and agent identity controls, which map to the verification, operations and agent permission gaps above. It supports a program. It does not replace one.

How does this link to NIST?

NIST's Secure Software Development Framework and its generative AI profile, SP 800-218A, extend secure development practices to AI-related work. The NIST AI Risk Management Framework adds Govern, Map, Measure and Manage functions. If a customer asks for your approach, mapping to these gives a recognised answer.

How to check or apply this

1. Run a one-day SAMM style self-assessment and score the five functions.
2. Pick the lowest two and set a ninety day improvement goal each.
3. Add blocking gates for known exploited vulnerabilities and exposed secrets.
4. Write and publish an AI tool policy.
5. Update threat models to include model inputs and agent tools.
6. Track pull request size and review time to catch overload early.

Quick check

Sources

• OWASP SAMM, Software Assurance Maturity Model
• OWASP Top 10 for LLM Applications, OWASP GenAI Security Project
• NIST SP 800-218, Secure Software Development Framework, NIST
• SP 800-218A, SSDF Community Profile for Generative AI, NIST
• AI Risk Management Framework, NIST

Last reviewed: 18 October 2026.