Open Policy Agent (OPA)

Describe Open Policy Agent (OPA) by the failure it prevents, not by its feature list. It belongs to policy as code and runs at the deploy stage of delivery, which matters because the cost of fixing a defect rises sharply the later it is found — the core argument behind shifting security left.

Strong answers cover the Rego language, OPA as a general decision engine vs Gatekeeper for Kubernetes admission, ConstraintTemplates and Constraints, policy unit testing with opa test, dry-run/audit mode before enforcing, and conftest for validating IaC before apply. Interviewers also listen for the operational side: who owns the rules, how findings reach the developer who introduced them, how false positives get suppressed without silently disabling coverage, and what the break-glass path is when a fix genuinely cannot ship in time.

Tie it to the wider toolchain. Security scanning is only useful when its output is actionable, deduplicated across tools, and attached to a specific commit, image digest or resource — otherwise teams learn to ignore it.

The usual failure mode is a security gate so slow or noisy that teams bypass it. For Open Policy Agent (OPA), the practical levers are: scan only what changed on pull requests and keep the exhaustive scan for a nightly or pre-release run; cache whatever database or ruleset the tool downloads so CI is not re-fetching it every job; and parallelise across services rather than scanning a monorepo serially.

Severity policy matters as much as speed. Blocking on every finding trains people to add blanket suppressions. A workable policy blocks on Critical (and often High) with a documented, time-boxed exception process, while everything lower is tracked as a ticket rather than a hard stop.

For this tool specifically, tune the Rego language, OPA as a general decision engine vs Gatekeeper for Kubernetes admission, ConstraintTemplates and Constraints, policy unit testing with opa test, dry-run/audit mode before enforcing, and conftest for validating IaC before apply. Finally, make the gate honest: an admission webhook denying a Pod that runs as root or lacks resource limits. A gate that warns but never blocks is documentation, not a control.

Every scanner produces findings that are not real risks in context — an unreachable code path, a vulnerable function the app never calls, a test fixture that looks like a credential. Answer this by describing a triage flow rather than claiming the tool is accurate.

First, separate "is this a true finding" from "is it exploitable here". Reachability, network exposure, and whether the affected component handles untrusted input all matter more than the CVSS number alone. Second, use the tool's own suppression path for Open Policy Agent (OPA) rather than deleting the check — the Rego language is usually where this is configured. Every suppression should carry an owner, a justification and an expiry date, and should be reviewed like code.

Third, baseline. Adopting a scanner on a mature codebase produces a wall of findings; freezing that backlog and blocking only on newly introduced issues is what makes adoption survivable. Then burn the backlog down deliberately.

Start with verification: a Critical rating is a claim, not a fact. Confirm the affected version or resource is actually deployed and that the vulnerable path is reachable from untrusted input.

Then contain according to the finding class. A leaked credential is rotated first and only then removed from history — deleting the commit does nothing if the key is already public. A vulnerable dependency is patched or the affected feature disabled. A compromised host is isolated for forensics before termination, since terminating destroys the evidence. A misconfigured policy is corrected and audited for other resources sharing it.

Close the loop in the pipeline. The point of DevSecOps is that the same class of issue should be caught earlier next time: add or tighten the rule, lower the threshold, or add a policy at admission. For Open Policy Agent (OPA), that usually means an admission webhook denying a Pod that runs as root or lacks resource limits.

Finish with the blameless review — how long the exposure lasted, why the existing gates missed it, and what detection you added.

Interviewers use "compare X to Y" to check you understand tradeoffs rather than tool names. For Open Policy Agent (OPA), the honest comparison is Kyverno for Kubernetes — OPA/Gatekeeper is more powerful and works beyond k8s (APIs, Terraform, CI) but requires learning Rego; Kyverno is k8s-only YAML.

Choose on concrete axes: what stage it runs at (deploy), how it integrates with your existing CI/CD and registries, how noisy it is out of the box, whether it is open-source or a paid managed service, and whether it covers the platforms you actually run. Coverage overlap matters too — running two scanners of the same class mostly multiplies triage work unless one meaningfully catches what the other misses.

Strong answers avoid "X is just better." They say when each wins: pick the tool whose stage, ecosystem fit and signal-to-noise match your constraints, and note where you would run something else alongside it.

Operating a scanner is not "turn it on" — it needs the same observability as any production system. For Open Policy Agent (OPA), the signals that matter are policy evaluation latency at admission, count of denied vs audited violations, and policy unit-test coverage.

Trend them. Rising Critical counts on a stable codebase usually means a base image or dependency has aged, not that developers got careless. A rising false-positive rate is an early warning that the gate is about to lose the team's trust. Scan duration matters because a slow gate is a gate people route around.

Feed these into a dashboard alongside remediation SLAs so security posture is visible to the people who own it, not buried in a tool nobody logs into.

Write policies in Rego and unit-test them with opa test against sample inputs so a bad policy fails in CI, not in the cluster. Roll out in dryrun/audit enforcementAction first to see what would be denied without blocking anyone, review the audit violations, then flip to deny. Validate infrastructure-as-code pre-merge with conftest so the same policies catch problems before apply, not just at admission.

Do start every constraint in audit mode and graduate to enforce after reviewing real violations; do not ship a deny policy untested, because a mistake blocks every deployment cluster-wide.

The anti-pattern usually comes from treating the tool as a checkbox rather than a control people have to live with. A gate that is too loud gets bypassed; a gate that blocks on things nobody can fix gets disabled; a suppression with no expiry becomes a permanent blind spot. The best practice above is the version that survives contact with a real team under delivery pressure.

Generally: make the signal trustworthy (low noise), make the required action clear and achievable, give an honest escape hatch for genuine exceptions, and review suppressions and thresholds on a schedule so the configuration does not quietly rot into "always green, never useful."