AI AppSec agents are here: a practical triage → patch workflow (without chaos)

Vulnerability management has a nasty failure mode: once the backlog gets big enough, everything becomes “later” — until it becomes an incident. AI security agents promise a shortcut: scan code, validate findings, propose patches.

The opportunity is real. The risk is also real: noisy findings, unsafe patches, or changes that aren’t reviewable. Here’s a workflow that keeps the speed, but makes the output auditable and shippable.

What’s changing in 2026

• Security agents are moving beyond “scan”. They’re starting to validate and patch with repo context, tests, and tool use.
• AppSec is converging with software delivery. The best security workflow is the one that lands as a clean PR with tests and a clear blast radius.
• Evaluation is becoming mandatory. If an agent writes code, you need repeatable checks (and regression tests) the same way you would for any developer tool.

The workflow: triage → validate → patch → ship

Step 1) Constrain scope (one repo + one class of issues)

Start narrow. Pick a single repo and a single class of issues (e.g. dependency vulns, auth bugs, SSRF). This makes it possible to measure quality and avoid “AI touched everything”.

Step 2) Convert findings into a structured case file

Don’t pass around screenshots and Slack paste. For each finding, the agent should produce a short structured object:

• finding_id, severity, component, paths
• exploitability (with assumptions)
• repro_steps or proof (tests, logs, links)
• recommended_fix + tradeoffs
• confidence (low/med/high)

Step 3) Use confidence tiers to control what the agent can do

• Low confidence: gather context, propose hypotheses, request a human.
• Medium confidence: open a draft PR, add tests, run linters (no merge).
• High confidence: open a PR with a full explanation + rollback notes (still no merge without review).

Step 4) Make “patches” a product (tests + rollback + changelog)

A patch that compiles isn’t a patch you can ship. Require the agent’s PR to include:

• tests that fail before / pass after (when feasible)
• a short threat model note (“what attack does this stop?”)
• rollout notes (feature flag, config toggle, or safe revert)
• a plain-English summary for non-security reviewers

Step 5) Add eval gates (and keep them forever)

Every time you change the agent’s prompt, tools, model, or permissions: re-run your evaluation suite. At minimum, keep:

• a set of historical vulnerabilities from your own repo (sanitised if needed)
• prompt-injection test cases (because the agent reads untrusted text)
• false-positive tests (to control noise)

The “SMB version” of AppSec maturity

You don’t need a giant security program to get value. The minimum viable version is:

• one scoped repo
• one weekly queue review
• draft PRs only (no auto-merge)
• logs of what the agent read and changed

Where Workflow ADL fits

We build safe, auditable AI workflows for real operations. If you want an AI-assisted AppSec pipeline (triage + draft PRs + eval gates + approval lanes) integrated with your existing CI/CD and ticketing, book a consult.

Freshness (RSS): OpenAI: Codex Security (research preview), OpenAI: acquiring Promptfoo, OpenAI: Improving instruction hierarchy in frontier LLMs.