Support teams inherit infrastructure pipelines they did not design, and the first outage is usually the first time anyone reads them closely. If your team looks after customer environments, a short list of weak spots will save you more pain than a long policy document. These eight come from the Ops Atlas entry on InfraOps, which collects them under anti-patterns and continuous-integration traps. Work down the list in order; each item takes minutes to check and costs nothing to fix.
1. Preview credentials that can read everything
The role that runs a preview is often described as read-only, and teams relax about it. The Atlas points out that it still sees every secret stored in the state, and that previews execute provider and data-source code. Handle preview credentials almost as carefully as the ones that apply changes: short-lived tokens, narrow scope and a log of every use. In a support setting, where one pipeline may serve several customers, this single change shrinks the damage from a leaked token considerably.
2. Any branch author can use the preview role
On GitHub, a pull request builds from the merge commit, including edits to the workflow file. That means a person who can open a pull request can alter what the job does with the preview role. Review who may push branches, and consider limiting the role to specific, reviewed workflow files.
3. The apply role trusts too much
The identity used for real changes should accept requests only from a protected production environment, not from a whole repository. Also check how your plan tier handles required reviewers: the Atlas notes that on GitHub Free, Pro and Team they work only in public repositories. For private repositories you may need a different gate, so confirm it works rather than assuming.
4. Plan files travelling between jobs
A saved plan contains state and variables. If your pipeline uploads it as a build artifact, anyone who can read the repository may be able to download it. Keep planning and applying in the same job where possible, or protect the artifact as if it were a credential. Plans also go off quickly. After one has been applied, a second plan drawn up against the earlier state no longer matches and cannot be trusted.
5. Laptops with apply rights
Every engineer able to apply from a personal machine is a path around your review process. During an incident the temptation is strong, which is exactly when mistakes happen. Remove the access, and make the pipeline quick enough that nobody feels they need a shortcut. If a standard change takes an hour to get through review, people will find a way around it, so measure the wait and trim it. Support engineers are often the ones with the broadest rights, which makes this item matter more for you than for most teams.
6. Targeting as a habit
The option that restricts a run to one resource is, in the tool’s own documentation, meant for exceptional circumstances only. If your support runbooks include it as a normal step, you are skipping part of the dependency graph every time. Replace it with a smaller state or a proper change. A good test is to ask how often the option appears in your shell history; if the answer is weekly, the real problem is the size of the state, which the Atlas, quoting Google Cloud, does not want to see above 100 items, with a few dozen as the ideal.
7. Breaking locks you did not take
When a run crashes, a lock can be left behind, and force-unlocking is tempting. Don’t, unless you have confirmed that the holder is gone. Never release a lock that someone else took; two writers on one state is how records get corrupted. Add a line to the runbook that says who to ask.
8. Expecting an undo button
Failed applies do not roll back. Some resources have changed and others have not, and the right response is a fresh plan that moves forward. Rehearse this once in a test account so the first real attempt is not also the first time anybody tries it. Write down what you saw, including how long it took, because that figure becomes the honest answer when a customer asks how quickly you can recover from a half-finished change.
A quick scoring sheet
Give each item a mark: closed, partly closed or open. Anything open on items one to three goes to the top of the queue, because those concern credentials. Items four to eight are about habits and runbooks, and they can be fixed one sprint at a time. Record the date, then repeat the check next quarter. Trends are more persuasive to a customer than a single audit.
Where this fits in support work
Managed-service teams carry the extra burden of explaining findings to customers in plain terms. Our posts on why operations engineers benefit from learning code-driven infrastructure and reliability habits and core principles behind dependable software delivery offer language for those conversations.
For the full list of traps, including the immutable subject format on GitHub for repositories created after July 2026, read the Ops Atlas page on InfraOps. It also contains a section on maturity levels if you need to explain where a customer sits today.
One last tip: print this list and pin it beside the on-call rota. The traps are not exotic, they are ordinary shortcuts, and a visible checklist is the cheapest control you can buy.