Hardening

Hardening

Assume input is hostile and the network is not your friend. Most vulnerabilities come from trusting something you shouldn't: user input, a token, a default config, an upstream response, or text stuffed into an LLM prompt. Validate at boundaries, authorize every action, protect secrets, and fail closed when something is wrong.

Security is not a final scan — it's built into every change that crosses a trust boundary. Retrofitting after ship is slower and always misses something.

Run the security checklist alongside this process for a quick pass. For supply-chain and dependency CVEs, pair with [[dependency-hygiene]]. For review workflow, see [[review-gate]] and the security-auditor agent. For logging without leaking data, see [[observability]].

When to Use

• Any input crosses a trust boundary — HTTP params, JSON bodies, headers, cookies, uploads, webhooks
• Authentication, authorization, sessions, JWTs, or API keys
• Storing, transmitting, or logging sensitive or personal data (PII, payment, health)
• Building or reviewing APIs, forms, admin tools, or multi-tenant features
• LLM features with user or external content in prompts ([[llm-feature-engineering]])
• Configuring production — TLS, headers, CORS, rate limits, secrets
• Security-focused PR review or pre-launch pass ([[launch-readiness]])

Skip as the primary skill for pure internal refactors with no boundary change — still scan if the diff touches auth, validation, or data access paths.

Process

Work in order. Map boundaries before diving into individual lines.

1. Map trust boundaries

List where untrusted data enters and where privileged actions happen:

Boundary	Untrusted inputs	Privileged actions
Public API	Query, body, headers	Read/write any user's data
Admin UI	Same + file uploads	Bulk export, role changes
Webhook handler	Raw body, signature header	Trigger internal jobs
Background job	Queue message, S3 key	DB writes, outbound calls
LLM pipeline	User text, retrieved docs	Tool calls, SQL generation
Browser	DOM, localStorage	Usually server must re-verify

Every arrow from untrusted → trusted is where controls must live. Client-side checks are UX, not security — repeat validation and authorization server-side.

2. Validate input at the boundary — allowlist

Validate as soon as data enters your code — type, range, format, size — before it reaches SQL, shell, templates, or business logic.

Allowlist what's acceptable; don't blocklist "bad" strings (attackers invent new ones).

Good: email must match RFC-ish pattern; age 0–120; enum status in {pending, paid}
Bad:  strip '<script>' and hope

Size limits on everything — body size, array length, string length, upload bytes. Unbounded input is a DoS vector.

Types — parse JSON/schema strictly; reject unknown fields when the contract is fixed ([[interface-design]]).

Files — verify type and size; store outside web root; generate safe names — never use user path as filesystem path (see uploads below).

Reject invalid input with a generic client message; log details server-side without secrets.

3. Never concatenate untrusted data into interpreters

Injection happens when untrusted data becomes syntax in SQL, shell, HTML, URLs, or templates.

Risk	Unsafe	Safe
SQL	"SELECT ... WHERE id = " + id	Parameterized queries / prepared statements
Shell	os.system("convert " + path)	Argument array APIs; no shell
HTML/XSS	Template with unescaped user HTML	Auto-escape templates; CSP
NoSQL	{$where: userInput}	Typed queries; operator allowlist
LDAP/XML	String concat	Library APIs with parameters

Output encoding matters at the sink — HTML context ≠ JS context ≠ URL context. Encode for where data lands.

For redirects and links: allowlist destinations or map IDs to URLs — don't redirect(userSuppliedUrl) (open redirect).

4. Authorize every protected action — server-side

Authentication (who are you?) ≠ authorization (may you do this to this object?).

For every endpoint and mutation:

1. Is the caller authenticated (if required)?
2. Is the caller allowed to perform this action?
3. On this specific resource (row, file, tenant)?

IDOR (insecure direct object reference) — classic bug:

GET /api/orders/123   — does 123 belong to this user/tenant?
DELETE /files/{id}    — verify ownership before delete

Check authorization in the handler or service layer — not only hiding buttons in the UI.

Multi-tenant: scope every query by tenant_id (or equivalent) from the authenticated session — never from a client-supplied field without verification ([[data-modeling]]).

Least privilege — roles, service accounts, and DB credentials get minimum permissions needed.

5. Sessions, tokens, and API keys

• Use standard flows — OAuth2/OIDC, session cookies with HttpOnly, Secure, SameSite
• Short-lived access tokens; refresh rotation where appropriate
• Invalidate on logout, password change, and compromise
• Don't store secrets in localStorage if avoidable — XSS steals them
• API keys: scoped, rotatable, hashed at rest, never logged
• Don't roll your own session tokens or password hashing — use vetted libraries (bcrypt/argon2, framework session stores)

Verify webhook signatures (HMAC, timestamp) before trusting body — treat unsigned webhooks as untrusted input.

6. Secrets and sensitive data

Never in:

• Source code, git history, screenshots
• Client bundles or mobile apps (assume extracted)
• Logs, metrics labels, error messages, support tickets auto-filled from stack traces
• URLs query params

Do:

• Load from secrets manager / env at runtime ([[git-flow]] — scan diffs)
• Rotate immediately if leaked — removal from git is not enough
• Scrub PII and secrets from logs and traces ([[observability]])
• Encrypt in transit (TLS everywhere) and at rest where required (DB, backups, disks)

Classify data — what's PII, payment, health — and model retention/delete paths ([[data-modeling]], [[decision-docs]] for compliance rationale).

7. Fail closed with safe defaults

When auth, validation, or dependency checks fail — deny, don't degrade to open:

Wrong	Right
catch { return true }	Reject request; log internally
Missing role → full access	Missing role → deny
Config default auth=false	Default deny; explicit opt-in to expose
Error returns stack trace to client	Generic message; details in server log

Production defaults:

• TLS on; HSTS where appropriate
• Security headers — Content-Security-Policy, X-Content-Type-Options, X-Frame-Options or CSP frame-ancestors
• Restrictive CORS — explicit origins, not * with credentials
• Debug endpoints and admin panels disabled or gated in prod
• Rate limits and abuse protection on login, signup, expensive APIs, and LLM endpoints

8. High-risk patterns — check explicitly

SSRF — user-supplied URLs fetched server-side:

• Allowlist domains; block metadata IPs (169.254.x, internal ranges)
• Don't pass raw URL to HTTP client from user input

File upload / path traversal

• Never open(userSuppliedPath) — map to internal ID or sanitized name
• Store outside web root; serve via controlled handler with authz check

Deserialization — don't deserialize untrusted data into rich objects (Java, Python pickle, etc.)

Mass assignment — don't bind request JSON directly to DB models without field allowlist

Caching — personalized or auth responses not cached under global keys ([[caching-strategy]])

LLM / AI ([[llm-feature-engineering]])

• Untrusted text in prompts = prompt injection surface
• Never eval model output; validate structured output against schema
• Tool calls from model: allowlist tools and arguments; authz as if user called API
• Cap input size; log prompts without PII where possible

9. Dependencies and known vulnerabilities

A vulnerable library is your vulnerability ([[dependency-hygiene]]):

• Pin and audit in CI
• Patch critical/high on production deps promptly
• Vet new packages before adopt

Don't disable security features of frameworks to "make it work."

10. Verify — don't assume secure

Before merge or launch:

1. Run the security checklist.
2. Trace one untrusted input through to storage/output — did validation and authz hold?
3. Try horizontal access — user A's token on user B's resource ID.
4. Trigger error paths — do responses leak internals?
5. Grep for secrets — git log, trufflehog, IDE patterns.
6. Confirm dependency audit clean or exceptions documented.

For significant surface changes, use security-auditor persona or peer review with attack scenarios.

11. Scenario playbooks

New public API endpoint

Boundary map → schema validation → authz on object → parameterized queries → rate limit → checklist.

Login / signup

Rate limit, credential stuffing protection, secure cookies, generic errors ("invalid credentials"), no user enumeration via error messages.

Admin or bulk export

Strong authz role check, audit log who exported what, limit row count, no IDOR on export filters.

Webhook receiver

Verify signature + timestamp → parse JSON → treat as untrusted → idempotent handler ([[resilience]]).

Multi-tenant feature

Tenant from session only; every query scoped; test cross-tenant access explicitly.

LLM feature with tools

Allowlist tools; validate tool args; same authz as direct API; block instruction override in user text where possible (delimiter patterns, separation — not foolproof, layer defenses).

Security regression in PR

Diff trust boundaries → focus review on changed edges → checklist → targeted tests for authz bypass.

Common Rationalizations

• "Our users wouldn't do that." — Attackers aren't your users; bots scan everything.
• "It's behind auth." — Authenticated users still attack IDs they shouldn't reach; check authz.
• "We'll add validation later." — Exploitable the moment it ships.
• "It's an internal service." — Internal boundaries get crossed; defense in depth.
• "We validate on the client." — Client is attacker-controlled; server must enforce.
• "CORS is fine — we use JWT." — Misconfigured CORS still matters; defense in layers.
• "It's only a low-severity CVE." — Chain lows into highs; track and time-box fixes.
• "Security headers break our embed." — Fix embed properly; don't drop CSP globally.
• "The LLM won't follow malicious instructions." — Assume it will; constrain tools and output.

Red Flags

• User input concatenated into SQL, shell, HTML, or template strings
• GET handler changes state without authz check
• Login checked but not permission on the specific resource (IDOR)
• tenant_id or userId taken from request body without verification
• Secrets in repo, config, client bundle, or logs
• Stack traces, internal paths, or tokens in API responses
• catch (e) { } or default-allow on auth failures
• Access-Control-Allow-Origin: * with credentials
• No rate limiting on auth or expensive endpoints
• User-controlled URL fetched server-side without SSRF controls
• File path built from user input
• eval, pickle.loads, or equivalent on untrusted data
• Model output executed or parsed without schema validation
• Outdated dependencies with known critical/high advisories
• Security checklist skipped for "small" API change

Verification

• Trust boundaries mapped; server-side enforcement on every edge
• Input validated at boundary by allowlist — type, range, format, size
• No injection sinks — parameterized queries, safe encoders, no concat
• Authorization checked per action and per object; tenant scoped ([[data-modeling]])
• Sessions/tokens standard, expiring, invalidated on logout; webhooks verified
• No secrets in code, logs, or client; PII scrubbed from telemetry ([[observability]])
• Fail closed — deny default, safe errors, TLS, headers, restrictive CORS
• SSRF, upload, deserialization, mass-assignment risks addressed if applicable
• LLM paths: no blind trust of output; tools allowlisted; injection layered ([[llm-feature-engineering]])
• Dependencies audited; highs addressed ([[dependency-hygiene]])
• Security checklist completed for the change