Security baseline

Most applications are not insecure because somebody wrote a bad line. They are insecure because a whole domain was never considered, and nothing about the code says so. This is the list of domains, what makes each one real rather than declared, and the order to take them in.

The rule: a control you cannot prove is a control you do not have — and a control that only new code respects is half a control.

---

1. Three things that invalidate defences you already paid for

Before any list, check these. Each one silently cancels controls that appear to be in place.

1. The origin must be reachable only through the edge. If the application answers requests that did not come through the CDN, WAF or proxy, then every edge rule — bot management, rate limits, IP reputation, the trusted client-address header — is an opt-in the attacker declines. Every control that trusts a header set by the edge is only as true as this.
2. After any credential or key exposure, rotate and then verify what the rotation broke. The application signing key usually underwrites sessions, cookies, signed URLs and stored tokens at once; rotating it is correct and it invalidates all four. Rotating without checking them is how a fix becomes an outage; not rotating is how an incident stays open.
3. A language with no static analysis in the pipeline has no automated floor. If the main language of the codebase is not covered by a scanner, every finding in it depends on someone noticing. Know which languages your pipeline actually covers — the default configuration of most tools does not cover all of them.

2. The domains

Walk them in this order the first time. references/controls.md has the individual controls for each, with what makes them real; read the section for the domain you are working on, not the whole file.

#	Domain	The question it answers
1	Access control and authorisation	Can one identity reach another's data or actions?
2	Authentication and identity	Can somebody become a user they are not? See padosoft-auth-hardening
3	Injection	Can input become code, a query, a command or markup?
4	Cryptography and secrets	Where do the secrets live, who can read them, when do they change?
5	Configuration	Is a safe default actually the default, everywhere, in every environment?
6	Software supply chain	Do you know what you ship, and would you notice if it changed?
7	Build and delivery integrity	Can somebody put code into production without going through the gate?
8	Logging, detection and alerting	If it happens tonight, who finds out, and how?
9	Exceptional conditions	When something fails, does it fail closed?
10	Rate limiting and anti-automation	What stops a script doing this ten thousand times?
11	Business-logic abuse	What happens if somebody uses the feature exactly as designed, maliciously?
12	Data protection and privacy	What do you hold, for how long, and who else gets it?
13	Uploads and files	Can a file become an execution path?
14	API surface	Which endpoints exist, and does each one carry its own authorisation?
15	Browser surface	What runs on your origin, and who put it there?
16	Mobile surface	What did you ship inside a bundle that anybody can open?
17	Infrastructure and edge	What is reachable, with which privileges, and what happens when it is lost?
18	Process and governance	What keeps all of the above true in six months?
19	AI and model surface	The model is a caller with unusual reach — see padosoft-agent-host-boundaries

Domains 1, 3, 9, 13 and 14 are where a review of the code finds things. Domains 5, 6, 7, 17 and 18 are where an assessment of the system finds things, and they are the ones that get skipped because no diff contains them.

3. Reading a status honestly

The whole value of an assessment is in refusing to overstate it. Four statuses, and the discipline is in the last two:

	Means
Covered	there is a control, and a gate or a rule applies it to new code
Native	the framework provides it unless somebody disables it — the risk is deactivation, not absence
Partial	it exists and does not cover the whole surface. Write what is missing, not just "partial"
Not inspected	nobody looked. Write how to look, so the next person can. This is an honest state

Two traps that turn an assessment into fiction:

• "Covered by a rule" does not mean the existing codebase complies. A rule is applied in review on new code. Code written before the rule existed has never been checked against it. Both facts are fine; conflating them is not.
• An identifier quoted in a comment is not a control. Code that says it implements a security rule, with no rule behind it, means the control exists today and nothing prevents the next change from removing it. The inverse of the previous trap: there, old code is unverified; here, future code is unguarded.

4. The rules that recur in every domain

These are not domain-specific. They decide whether a control is real.

• A setting is not a boundary. An authorisation or security boundary must not be switchable at runtime by anybody who can reach a settings screen. If it must be configurable, the configuration cannot reach off, and the floor is in code.
• An empty configuration value must not mean "disabled". The absent case is the common case: a fresh environment, a tenant nobody configured, a variable with a typo. Absent means the secure default, and disabling takes an explicit, visible value.
• A control that fails open needs a written reason and a second control behind it. Failing open is sometimes right — refusing every login because a third-party verifier is unreachable is worse than the friction it was buying. It is only acceptable while something else carries the weight, and that dependency is written down so that removing the other control re-opens the question.
• A control that fails closed needs an escape. A guard that denies on its own internal error is correct and will one day deny everybody; know which of the two you chose, and why.
• Half a switch is worse than none. A toggle that disables the server-side check while the client keeps behaving as if the control were on gives the impression of a decision nobody actually made.
• Twin endpoints must be protected identically. Where two routes reach the same resource — an older and a newer one, a page and its asynchronous counterpart, a web form and an API — the one built last usually has the controls, and the one built first is still wired.
• A gate that is born red and left red gets switched off within a week, and then it protects nothing at all. Introduce a strict check on new work first, and bring the backlog in behind it.
• Remove an endpoint rather than gating it. A route that exists in order to be refused is a route that one configuration change makes reachable.

5. Turning a finding into a control

A closed finding is worth very little; the class it belongs to is worth a lot.

1. Fix it in a small change, with a test that fails without the fix.
2. Name the class. Not "this endpoint leaked the model name" but "error messages must not carry internal identifiers".
3. Write the rule, and put the check where it will run without anyone remembering — a validator, a grep in review, a test, a gate.
4. Label the change, so the set of security work is answerable in one query when an auditor, a customer or your own team asks what was done and what is still open. A security fix that stays open is an exposure window, not a backlog item.
5. Find the twins. The same class almost always exists somewhere else in the codebase.

6. Where to start, on a new application

In order, because each one makes the next cheaper:

1. Authorisation derived from the authenticated identity, never from a parameter — padosoft-tenant-isolation.
2. Errors that carry nothing internal, and logs that carry nothing personal — padosoft-logging-discipline.
3. Rate limiting on everything that can be attempted repeatedly — padosoft-auth-hardening.
4. Secure defaults for headers, cookies and the environment gate — padosoft-environment-gating.
5. Dependency and secret scanning in the pipeline, blocking on the diff — padosoft-ci-workflow-gates.
6. One audit trail whose integrity can be shown — padosoft-evidence-boundaries.
7. Only then the long tail in references/controls.md.

---

Gotchas

• A penetration test is a sample, not a coverage measure. Closing its findings raises the floor by exactly the height of that sample.
• The report that arrives too often is not read. A monthly access review gets attention; a weekly one becomes a filter rule — see padosoft-failure-visibility.
• The control everybody forgets is the one for privileges granted outside the role system. Reviewing roles does not show them, and they are the channel through which access grows silently.
• A configuration flag and an environment name are two gates, and the dangerous arrangement is the one where either alone turns a convenience on.
• Owning a scanner is not running it, and running it is not reading it. The finding with no owner and no date is the same as no finding.
• An exception in an allow-list outlives the thing that justified it. When that name comes back, the exception authorises it silently. Expire them, or report the orphans.

Checklist

• Origin reachable only through the edge; every header-trusting control depends on it
• Keys rotated after any exposure, and what the rotation invalidated has been verified
• Static analysis covers the main language of the codebase
• All 19 domains have a status, and not inspected says how to inspect
• No status claims coverage of the existing codebase when the rule only applies to new code
• No security boundary governed by a runtime setting; no empty value meaning "off"
• Every fail-open documented with the control that carries the weight
• Twin endpoints compared side by side
• Each closed finding has a named class, a rule and a check that runs by itself
• Security changes labelled so the whole set is answerable in one query

Final report

Scope: <application / service>   ·   assessed against: <standards>
Domains: covered <n> · native <n> · partial <n> · not inspected <n>
Invalidators: edge-only origin <…> · key rotation <…> · static analysis <…>
Top gaps, in the order to take them: 1. <…>  2. <…>  3. <…>
Rules added from findings: <which, and where the check runs>
Stated honestly: <what is "partial" and why> · <what nobody looked at>