Safety management rests on one habit: replacing comfortable assumptions with checked facts. Hazard reports, risk assessments and audits all exist because things that “everyone knows” turn out, now and then, to be wrong. Access control in cloud systems collects its own set of comfortable assumptions, and the people who study it under the informal label IAMOps have catalogued the ones that cause the most damage. This piece tests six of them, one at a time, in the measured manner a safety office would expect.
A short note on the source
The myths below follow the reasoning in the Ops Atlas entry on IAMOps, which describes the word as an informal label and applies it to authorisation as code. Where a claim is attributed to the Atlas it reflects that reference; where I generalise, I say so.
Myth 1: “Open the doors wide today and narrow things down afterwards”
The belief: getting work moving matters more, and cleaning up can wait for a calmer week.
The reality: the Atlas points out that later never arrives on schedule, and when it does, the activity log written under broad rights seems to prove the rights were needed. A safety professional would recognise the pattern as a normalised deviation: a temporary exception becomes the standard. The more reliable approach is to begin narrow and widen a specific permission with a recorded reason.
Myth 2: “A ready-made permission bundle is already least privilege”
The belief: the provider’s pre-built roles were designed by experts, so using one means you are covered.
The reality: according to the Atlas, a leading cloud provider’s own documentation cautions that its prebuilt policies may fall short of least privilege. They are a convenient starting point and a poor finish line. Compare a generic emergency checklist from a manufacturer: it is the base document, and an operator still tailors it to its own aircraft and operation.
Myth 3: “An organisation-wide allow rule gives everyone access”
The belief: if the top-level policy says allow, people can proceed.
The reality: the Atlas explains that organisation-level control policies only set the outer limit of what is possible. They never grant access by themselves, and a clear denial wins wherever one applies. A useful parallel is airspace: a controlled zone being open does not give any pilot clearance to enter it.
Myth 4: “A condition on the repository name makes the role safe”
The belief: restricting trust to anything under one organisation’s namespace is specific enough.
The reality: the Atlas notes that a wildcard pattern such as the organisation’s name followed by a star admits every repository the organisation owns, including old or experimental ones. Exact, per-environment identifiers are the sturdier choice. In safety terms, “any member of the department” is not an authorisation; a named person for a named task is.
Myth 5: “If a rule looks right, it works”
The belief: reading a policy and finding it sensible is the same as verifying it.
The reality: the Atlas documents rule-engine behaviour that surprises reviewers. A negated test, such as “not equal to”, is satisfied if the value is simply absent, so a guardrail can wave through exactly the requests it targets. And one policy language skips a rule that fails with an error, so a broken prohibition fails open. The lesson is the same one a safety audit teaches: inspect by test, not by eyeball. Validate rules against a schema and run them in a sandbox before trusting them.
Myth 6: “Tightening access always makes things safer and nobody pays”
The belief: the more restrictive, the better, with no downside.
The reality: restriction has costs, and the Atlas is candid about it. In one worked example, the time to get a legitimate permission rose from 36 to 48 hours after a cleanup, which the team considered acceptable. Over-tightening also breaks rare tasks, such as a quarterly job that looks idle in a short window of usage data. Good practice balances the two by checking with owners before removal and watching denied requests after each change.
What the evidence says about doing nothing
The cost of leaving these myths unchallenged shows up in public data. The Atlas cites an OWASP ranking published in 2025 that put broken access control first, and reported the flaw in every application that was tested. A cloud security survey from the same year found year-old keys still active on 59% of AWS user accounts and 40% of Entra applications. A January 2024 intrusion at a large software company, as the Atlas tells it, started from a forgotten test app whose rights had never been lowered.
Why myths survive in technical teams
Safety culture offers a useful explanation. Myths persist because each one saves time on a typical day and costs nothing visible until an unusual day arrives. Granting wide access is quick. Trusting a vendor bundle is comfortable. Reading a rule instead of testing it feels like diligence. Nobody is reckless; the incentives simply favour speed, and the bill comes later, usually to someone other than the person who made the shortcut.
The remedy is the same one that works in flight operations: make the safe path the easy path. Put policies where reviewers can see them, add automatic tests so checking costs no extra effort, and give people a clear, fast route to request exceptions. When the proper route is quick, fewer people invent their own.
How to run a myth review in your own team
Pick one myth per month and treat it like a hazard report. Write the belief in a single sentence, name the system where it might apply, and assign one person to gather evidence. The output should be a short note: confirmed, disproved or partly true, with a link to the proof. Notes accumulate into a record that new colleagues can read in an afternoon, and the exercise doubles as training for people who rarely see how access rules behave under pressure.
A measured response
- Inventory every identity and give each an owner.
- Compare permissions granted with permissions used for your widest roles; in the Atlas’s example, one role used 27 of 180 actions, or 15%.
- Move policy changes into reviewed, tested files.
- Hold a recurring review and record exceptions with end dates.
These steps map neatly onto a safety management cycle: identify hazards, assess risk, apply controls, monitor, and review. None of it demands exotic technology, only the discipline to treat assumptions as hypotheses. Test one myth in your own systems this month, record what you find, and let the evidence set the next step.
