See how machine learning is changing secure SDLC practices, from AI-assisted code review and finding triage to new risks like model supply chain attacks.
TL;DR
- AI DevSecOps cuts both ways. Agentic workflows are now the fastest way to sort a pile of security findings, and AI systems are themselves a new class of attack surface.
- The defensive wins are real. Frontier LLMs have multiplied the number of findings faster than teams can work through them, widening the triage and remediation bottleneck instead of closing it. Agentic workflows, not just faster dashboards, are what close that gap now.
- Three risks are genuinely new: model supply chain tampering, training-data poisoning, and prompt injection reaching your applications in production.
- Maturity comes down to governance, not to how many AI features a team has switched on or built into their applications.
- Harness orchestrates scanners across the pipeline, triages findings with Triage Agent, remediates them with Remediation Agent, and enforces policy so security keeps pace with AI-speed delivery - with full audit trail from one end of the pipeline to the other.
What is AI DevSecOps?
AI DevSecOps is the practice of applying machine learning across the secure software development lifecycle, both to automate security work like scanning, triage, and remediation, and to govern the new risks that AI systems introduce into that same lifecycle. Traditional DevSecOps assumes the code, the dependencies, and the pipeline are the things you secure. AI DevSecOps adds three more assets to that list: the models, the prompts, and the data those models learn from.
- The defensive half. Models cluster and deduplicate scanner output, Triage Agent rank findings by false positive likelihood, exploitability, and reachability, and then Remediation Agent generates and validates the fix so the developer starts from a PR instead of a ticket.
- The risk half. Every model, prompt template, and training added to an application now needs provenance, policy, and monitoring, the same way a third-party library does.
- What hasn't changed. Security still has to run inside delivery rather than beside it, and a person still owns every exception.
Most writing about AI DevSecOps only tells you the good half. Models sort findings, cut false positives, draft patches. All true.
Here is the other half. In the same year your team started shipping AI-written code, your delivery process quietly picked up a set of dependencies nobody in AppSec has a playbook for: model weights pulled from a public hub, prompt templates checked into a repo, training data nobody diffused. AI in software development changed what you build with, so it changed what you have to secure.
How is AI in software development actually used in DevSecOps pipelines?
Three uses survive contact with a real pipeline.
Finding triage and deduplication. Run SAST, SCA, DAST, secrets, and container scanning on one commit and the same vulnerability shows up four ways. Models cluster those into one finding, drop duplicates, and rank the rest by whether the code path is reachable. It is the least glamorous of the four, and the one that pays for itself fastest. Harness Security Testing Agent does this across more than 40 commercial and open source scanners.
First-pass code review. Models flag injection patterns, hardcoded credentials, and unsafe deserialization before a human opens the diff. The pitch for shift left security AI is exactly this: scan earlier, and let the model absorb the noise that made early scanning unbearable last time.
Risk scoring for release decisions. CVSS alone says nothing about your blast radius; combined with reachability, service criticality, and deploy history it becomes a number a gate can act on.
Where does AI still fall short?
AI adds work as fast as it removes work. 53% of developers name reviewing AI-generated code for accuracy as a top source of friction (Harness, State of Engineering Excellence 2026). A tool that clears 400 findings and waves through one real exemption has cost more than it saved.
What new risks does AI introduce to the secure SDLC?
Three, and none maps cleanly onto a scanner you already own.
Model supply chain. A model pulled from a public hub is a binary artifact with thin provenance, in formats that can execute code on load. You would not pull an unsigned JAR into production, yet model weights arrive with no signature, no owner, and no entry in the software bill of materials. So treat them like any other dependency: pin the version, record its origin, scan it, list it.
Training and retrieval data poisoning. Corrupt what a model learns from, or retrieves at inference time, and you steer its output without touching application code. MITRE's ATLAS knowledge base catalogs these techniques against real machine learning systems, with case studies attached.
Prompt injection reaching your application. The one that matters most once your own product ships an agent - and that agent shipped through the same delivery pipeline you're trying to secure, so it inherits every gap upstream. A user-facing assistant that reads support tickets, browses the web, or calls internal tools will eventually process an instruction nobody authorized: a line buried in a document, a webpage, or a database record, telling it to exfiltrate data, call a tool it shouldn't, or ignore its own guardrails. OWASP ranks prompt injection first in its Top 10 for LLM applications, and exposure grows with the autonomy you hand the agent.
The risk nobody scanned for: 48% of teams are concerned about increased software vulnerabilities from AI coding assistants. Source: Harness, The State of AI in Software Engineering.
That exposure starts earlier than any pipeline scan. Code written inside an AI coding agent - Cursor, Devin, Claude Code, and similar tools - already exists before it ever reaches a commit, and most scanning still waits for that commit to happen. The vulnerability ships the moment the agent writes it, not the moment CI runs.
DevSecOps best practices and shift left security AI: what maturity looks like
Maturity here has little to do with how many AI features a team switched on. It tracks how much of the AI in the delivery process anyone can account for. Four stages, and most teams sit in the second.
The step that matters is tool-assisted to governed: an inventory of every model and prompt the delivery process depends on, an owner for each, policy that executes in the pipeline instead of a wiki, and an audit trail an assessor can read. NIST's AI Risk Management Framework is a sane spine, and it maps onto controls you already report against. The DevSecOps best practices you run today still apply, with a new class of assets to cover.
DORA's 2025 research puts AI's primary role as an amplifier of whatever an organization is already good and bad at. Governance decides which side gets amplified.
How Harness supports AI DevSecOps
The challenge
Security testing ran late long before AI arrived. Then AI multiplied the code moving through it, and now frontier LLMs doing the scanning have multiplied the findings on top of that. Findings stack up faster than any sprint can clear, scanners disagree about which are real, and whoever has to fix them wrote that code three weeks ago. 86% of engineering leaders and practitioners say security and compliance checks need to be more automated to meet delivery timelines (Harness, 2026 State of DevOps Modernization).
The approach
Harness is an AI platform for software delivery, and its security work runs inside the pipeline.
- Catch it at the source. Secure AI Coding integrates directly with the AI coding agents writing the code - Cursor, Devin, Claude Code, and others - to find and fix vulnerabilities in AI-generated code at the moment of generation, so nothing insecure makes it to your repo in the first place.
- Orchestrate, then deduplicate. Security Testing Orchestration normalizes results across scanners into one prioritized list, so a developer opens one finding instead of four.
- Triage automatically. Triage Agent ranks findings by false-positive likelihood, exploitability (EPSS), and reachability, so the queue that reaches a developer is already ordered by what actually matters.
- Remediate at speed, with humans in the loop. Remediation Agent generates a fix, checks that it does not break the build, and opens a PR with the validation detail attached. Developers review and approve the PR, without having to touch code.
- Govern the supply chain. Supply Chain Security generates and attests SBOMs and AI-BOMs, verifies SLSA provenance, and applies allow and deny policy to open source components and to the models, prompts, and agents your applications now depend on.
- Enforce as code. Policy as code, RBAC, and central exemption management turn a rule into a pipeline step that stops the build.
The outcome
Findings arrive in the pull request that caused them, with a proposed fix. Exceptions carry an owner and an expiry. Audit evidence becomes a query. None of that eliminates exposure, but it shortens the gap between a vulnerability appearing and its fix. A demo of secure CI/CD with Harness walks the pipeline end to end.
What does governed AI DevSecOps look like when it works?
How did Gameopedia cut QA maintenance by more than 40%?
Gameopedia's QA team was losing most of its week to test upkeep: brittle selectors, flaky reruns, new hires who needed a month to contribute. Moving test creation and maintenance onto Harness AI Test Automation cut QA maintenance by more than 40% and pulled that ramp-up from roughly 30 days to five or fewer.
“With Harness AI Test Automation, I would get back 2-3 hours daily, and I can use that time to be more productive and improve the tests for any edge case scenarios.”
Ajith KV, QA Engineering Lead, Gameopedia
Source: Gameopedia cuts QA maintenance 40% and speeds onboarding with Harness AI Test Automation
How did DigiCert drop new-region setup from weeks to minutes?
DigiCert secures digital infrastructure for much of the Fortune 100 and committed to moving from a 99.9% to a 99.99% uptime SLA. Swapping custom deployment scripts across Kubernetes, AWS, and on-premises environments for one governed pipeline is what made that consistency reachable, and standing up a new global region dropped from weeks to minutes.
“We weren't looking for a tool. We were looking for a foundation. Something that could support every team, every environment, every region we were moving into.”
Jared Daniels, Principal Software Engineer, DigiCert
Source: DigiCert achieves 99.99% reliability at global scale with Harness
Securing what AI brought in with it
The good half holds up. Machine learning is the best tool anyone has for making a pile of security findings legible. But the other half is what teams keep discovering late: models, prompts, and training data are dependencies now, and they need the provenance and policy already given to every library.
In the Harness State of AI in Software Engineering survey, 74% of respondents agreed that organizations failing to integrate AI safely and securely across the SDLC within a year will go the way of the dinosaurs. The work is unglamorous. Inventory, ownership, and policy that runs where the code runs.
