AI Writes Code Fast. Can You Trust What Gets Deployed? | Harness Blog

Based on the DevOps.com webinar "AI Writes Code Fast. Can You Trust What Gets Deployed?" presented by Harness, August 12, 2026.

AI has removed the bottleneck in writing software. Code that used to take days now takes minutes. But speed of creation and trustworthiness of what reaches production are two very different things. As AI increases the volume and velocity of code changes, the security model has to shift too: from reviewing code and periodically scanning artifacts to continuously proving that the exact software entering production can still be trusted.

The Artifact Is the New Attack Surface

Vulnerabilities are now discovered faster than most teams can fix them, and attackers have shifted their focus from source code to what actually gets deployed. The session walked through three real-world patterns of compromise:

• Input compromise (TanStack): a trusted dependency is compromised upstream a malicious package gets published, pulled into CI/CD builds, and shipped inside production artifacts.
• Build compromise (SolarWinds): the build system itself is compromised, malicious code is inserted, and a trusted update is distributed to every downstream customer.
• Runtime compromise (LiteLLM): a pipeline is modified to inject a backdoor, malicious code executes inside CI/CD, and secrets are exfiltrated.

These attacks expose the limitation of asking only, “Did we scan this?” The more important question is: Can I prove that the exact artifact I'm about to deploy is the artifact I intended to build, that it hasn't been compromised, and that it still satisfies my security policies right now?

Trust Has to Travel With the Artifact

Traditional scanners answer a point-in-time question. But a clean scan today doesn't mean a safe artifact next week, after a new CVE is disclosed against a dependency that was already approved.

The webinar's core argument was that security evidence needs to travel with the artifact itself. That includes its SBOM, provenance and attestations, vulnerability findings, policy evaluations, and the history of how it was built and promoted. Instead of security living in a separate report, the artifact carries the evidence needed to decide whether it can move forward.

That turns the artifact registry from a place where binaries are stored into a control point between build and production. Organizations can evaluate what enters the registry, continuously reassess what's already there as threats change, and enforce what is allowed to move into higher environments.

Detection Is Only Useful If You Can Enforce It

Continuous visibility is only half of the equation. When an artifact no longer meets policy, the system needs to be able to act on that information.

That means applying controls before artifacts enter the registry, such as vulnerability, license, age, and upstream trust policies, and controls after ingestion, such as quarantine, RBAC, promotion gates, and deployment policies. If a new vulnerability changes the risk of an artifact that was previously approved, that artifact can be identified and prevented from progressing rather than remaining trusted indefinitely.

The goal is a closed loop: detect a change in risk, evaluate it against policy, enforce the decision, and trigger remediation.

Where AI Agents Fit Into the Defense

Harness's own AI and agent stack was presented as a working example of what continuous verification looks like in practice:

• DevOps Agent — creates and edits pipelines in seconds, analyzes and fixes pipeline errors, and creates OPA policies personalized to your environment.
• AppSec Agent — automatically generates code fixes and PRs, surfaces contextual risk insights, defines policies to block risky builds and deploys, and monitors security posture in real time.
• Firewall Agent (on the roadmap) — scans the entire artifact registry against live threat feeds and proactively blocks or warns on new CVEs as they're disclosed, not just at scan time.
• Lifecycle Agent (on the roadmap) — gives full registry catalog visibility, detects orphaned, expired, and vulnerable artifacts, and triggers soft-delete and remediation workflows automatically.

From "Built" to "Trusted"

The shift the session pushed teams to make is moving from asking whether something was safe when it was built, to continuously re-evaluating whether it's still safe now that new threats have emerged. That means policy-driven promotion between environments, automated remediation triggered by newly discovered CVEs, and agentic response that acts on a threat the moment it's identified rather than waiting for the next scheduled scan.

The Takeaway for Security and Platform Teams

AI-generated code isn't the risk by itself; unverified trust in what gets deployed is. Organizations that have already solved the "inner loop" (writing code faster with AI) are discovering the outer loop,safely getting that code into production, is now the real bottleneck. Software delivery pipelines have effectively become security boundaries, and treating them that way, with continuous verification and artifact provenance built in, is what lets teams keep shipping fast without shipping blind.