Secure go

Secure Go Development

Write secure Go applications through idiomatic Go — explicit error handling, strict typing, a strong standard library, and built-in tooling (go vet, go test -race, govulncheck). Vulnerability is just a bug: code doing something the spec didn't ask for.

Go provides a strong starting position: no buffer overflows (GC, bounds-checked slices), strict static typing (no silent coercions), explicit error handling (no swallowed exceptions), a single static binary (smaller attack surface), and go.sum + checksum database guaranteeing dependency integrity out of the box. Go cryptographic packages passed an independent Trail of Bits audit in 2025 (commissioned by Google): 1 low-severity finding (fixed in Go 1.24), 5 informational.

Structure follows OWASP Top 10 2021 with 2025 mappings noted in each section.

---

1. Access Control — don't give more than needed

OWASP: A01:2021 / A01:2025 — Broken Access Control. SSRF (A10:2021) consolidated here.

Every data operation checks that the current user has the right to that operation on that data.

Code examples: references/access-and-data.md

Rules

• Deny access to all non-public resources by default.
• Authorization logic lives in middleware, reused across routes.
• Data is bound to its owner; ownership checked on every access (WHERE user_id = $1).
• Access rules covered by tests — they are business logic.

Recommended libraries

• Casbin — RBAC/ABAC as configuration. Adapters for Gin, Echo, Fiber, Chi.
• Oso — policy engine with declarative language Polar. For rules more complex than roles.

---

2. Data Protection — protect what shouldn't be public

OWASP: A02:2021 → A04:2025 — Cryptographic Failures.

Code examples: references/access-and-data.md — bcrypt/Argon2 hashing, env secrets, crypto/rand, API response structs.

Rules

• Don't store what you don't need. Delete or anonymize data once it served its purpose.
• Passwords: hash only (bcrypt/Argon2), never MD5/SHA-256 without salt.
• TLS everywhere data traverses between components.
• Secrets never live in source code or git.
• Tokens, keys, salt: only crypto/rand.

Recommended libraries & tools

• golang.org/x/crypto/bcrypt & golang.org/x/crypto/argon2 — password hashing. bcrypt simpler (manages salt and packs parameters into string), Argon2id adjustable in memory, preferred for new systems; bcrypt with cost 12+ remains acceptable per OWASP. Practical guidance from OWASP Password Storage Cheat Sheet: baseline profile m=19 MiB, t=2, p=1, keyLen=32 (minimum sufficient); heavier profile m=64 MiB, t=1, p=4 if hardware allows. Target ~0.5–1 second CPU per hash.
• crypto/rand — random values (salt, tokens, keys).
• crypto/subtle — ConstantTimeCompare for timing-safe hash comparisons.
• golang-jwt/jwt/v5 — JWT with mandatory algorithm verification.
• koanf — typed config from env, files, flags, Consul, etcd.
• env — minimal env→struct mapping via tags.
• viper — full-featured config combiner (env + yaml + consul + remote).
• SOPS — encrypt secrets in yaml/json config files (AWS KMS, GCP KMS, age, PGP). Safe in git.
• Vault — dynamic secret generation, rotation, access audit. Go client: hashicorp/vault/api.
• gitleaks — pre-commit hook scanning diffs for secret patterns.

---

3. Input Handling — trust only what is explicitly allowed

OWASP: A03:2021 → A05:2025 — Injection (XSS included in 2025).

Two universal rules: (1) never construct executable strings from raw user input; (2) validate on input, sanitize on output.

Code examples: references/input-and-design.md — parameterized SQL, sqlc, ORM raw queries, OS commands, html/template, validation.

Rules

• SQL — only through placeholders. fmt.Sprintf for SQL is a bug.
• OS commands — exec.Command with individual args, no sh -c.
• HTML — html/template, not text/template. Never wrap user data in template.HTML without sanitization.
• Validate on the server side, by whitelist (allowed chars, length, format).
• Sanitize context-dependently on output (HTML, URL, SQL — different contexts).

Recommended libraries & tools

• database/sql — parameterized queries out of the box.
• sqlc — typed Go from SQL. Injection structurally impossible.
• sqlx — database/sql extension with named queries.
• ent & GORM — ORMs with parameterization by default. Raw queries still need ? placeholders.
• html/template — stdlib, context-dependent auto-escaping.
• go-playground/validator — declarative validation via struct tags.
• ozzo-validation — code-based validation (no tags).
• bluemonday — HTML sanitization on a whitelist approach.

---

4. Secure Design — trust boundaries at design time

OWASP: A04:2021 → A06:2025 — Insecure Design.

Problems that can't be fixed at implementation because they're baked into the architecture.

Code examples: references/input-and-design.md — server-side business logic, race-condition protection, resource limits, rate limiting.

Rules

• For every feature, define who can do what and with what limits.
• Don't trust the client. Validate constraints on the server. Prices, discounts — server-side only.
• Resource consumption limits at the business-logic level.
• Each component gets exactly the permissions it needs.
• Use ready-made solutions for standard tasks (auth, hashing, sessions).
• Design habit: ask "What if someone does this 100,000 times? Substitutes someone else's ID? Passes a negative quantity?"

Recommended libraries & tools

• database/sql transactions — BeginTx + SELECT ... FOR UPDATE for atomic business operations.
• golang.org/x/time/rate — rate limiter from Go extended library.
• Go race detector (go test -race) — built-in.

---

5. Configuration — secure by default

OWASP: A05:2021 → A02:2025 — Security Misconfiguration. Moved to A02 in 2025.

Code examples: references/config-and-deps.md — production mode, error responses, security headers, file upload validation, server timeouts, Dockerfile.

Rules

• Default accounts/passwords — not in production.
• Detailed errors — only in dev. Client gets generic message + request_id.
• Strip everything unnecessary from distribution: test pages, framework docs, debug endpoints.
• Security headers configured once in middleware.
• http.Server always has ReadHeaderTimeout/ReadTimeout/WriteTimeout. TLS ≥ 1.2.
• Environment configuration is automated and reproducible (Dockerfile, docker-compose, Terraform).
• Docker: multi-stage builds, distroless runtime, non-root user (USER nonroot). Pin specific patch version + sha256 digest in production builds to prevent silently changed images on rebuild.
• Build flags: -trimpath strips paths; -ldflags="-s -w" strips debug info; CGO_ENABLED=0 for static binaries. Trade-off: -s -w removes symbol table and DWARF data — this shrinks binary and removes internal identifiers from dumps, but breaks profilers, debuggers, and symbolic stack traces. For production profiling (pprof), use only -trimpath.

Recommended libraries & tools

• secure — security headers middleware for Gin, Echo, Chi, net/http. One line: r.Use(secure.New(secure.Options{...}).Handler).
• Echo Secure middleware — built into github.com/labstack/echo/v4/middleware. Covers CSP, HSTS, X-Frame-Options, XSS-Protection, content-type-nosniff.
• http.DetectContentType — stdlib, MIME detection from first 512 bytes. Use io.ReadAtLeast (not io.ReadFull) — small files < 512 bytes are valid.

---

6. Dependencies — manage what you use

OWASP: A06:2021 → A03:2025 — Software Supply Chain Failures. Category expanded: beyond outdated dependencies, now explicitly covers supply chain risks — typosquatting, compromised maintainer accounts, registry package substitution.

Code examples & CI config: references/config-and-deps.md — govulncheck, CI workflow, version pinning, toolchain updates.

Rules

• Track dependencies (including transitive): go mod graph.
• govulncheck in CI — 5 lines in the workflow. Builds call-graph, checks reachability.
• go mod tidy — no unused dependencies.
• go mod verify — checksum integrity.
• Update dependencies regularly, but let new versions "settle."
• Use proven modules with active maintenance.
• Never run containers as root. USER nonroot in Dockerfile.
• Keep Go toolchain current (1.26 as of mid-2026).

Recommended libraries & tools

• govulncheck — official vulnerability checker from the Go team. Must-have.
• nancy — alternative scanner from Sonatype.
• Dependabot / Renovate — automated dependency update PRs.
• Go module proxy + checksum database (sum.golang.org) — built-in. go get verifies hashes transparently.

---

7. Authentication — ready-made solutions, correctly configured

OWASP: A07:2021 → A07:2025 — Authentication Failures.

Delegate at the highest possible level: identity platform > OIDC client > JWT library.

Code examples: references/auth-and-errors.md — JWT generation/parsing, per-IP rate limiter, breach password check, uniform error responses.

Rules

• Use ready-made libraries. Don't write your own authentication.
• JWT: always with exp, always with algorithm check. Prefer EdDSA/RS256 for new systems.
• Passwords: minimum length enforced.
• Session IDs: only crypto/rand.
• On logout: full session invalidation. For server-side sessions — delete the record. For JWT — denylist by jti or short exp + refresh token rotation with revocation. Without invalidation, "logout" only removes the cookie, but a stolen token keeps working until expiry.
• Rate limiting on login: by IP + by account.
• Uniform error responses on authentication failures. No user enumeration.

Recommended libraries & tools

• golang-jwt/jwt/v5 — JWT library (signing/parsing only). Always v5 with algorithm verification.
• coreos/go-oidc — OpenID Connect client. Delegate to Google, GitHub, Keycloak, Auth0.
• Ory Kratos, SuperTokens, ZITADEL — full identity platforms (login flow, MFA, password reset, audit log).
• golang.org/x/time/rate — per-IP rate limiting.
• ulule/limiter — rate limiter with Redis. Middleware for Gin, Echo, Chi.
• gorilla/sessions — server-side sessions (cookie, Redis, PostgreSQL, filesystem).
• mattevans/pwned-passwords — breach check via k-anonymity.

---

8. Error Handling — exceptions must not become holes

OWASP: A10:2025 — Mishandling of Exceptional Conditions (new in 2025).

Code examples: references/auth-and-errors.md — per-goroutine recover, transactional multi-step ops, recovery middleware, committed-flag cleanup, errgroup.

Rules

• Long-lived goroutines (background workers, queues): defer recover() at the isolation boundary. One-off goroutines don't need silencing — software bugs should crash loudly and be visible in tests.
• Multi-step business operations: transactions with defer tx.Rollback().
• Recovery middleware: standard for request goroutines, custom for own workers. Never leak stack traces to clients.
• Resources (files, connections): cleanup on errors via defer.
• panic — for unrecoverable situations only. Not for normal errors.

Recommended libraries & tools

• gin.Recovery() / echo.Recover() / fiber.Recover() — built-in recovery middleware for HTTP request goroutines.
• errgroup — goroutine coordination. One fails → others canceled via context.
• database/sql transactions — defer tx.Rollback() idiom.
• defer — resource cleanup on errors.

---

9. Data & Software Integrity — verify what you receive

OWASP: A08:2021 → A08:2025 — Software or Data Integrity Failures. In 2025 partially redistributed: supply chain moved to A03:2025, data/signature integrity verification remained as A08.

Code examples: references/integrity-and-logging.md — HMAC signing, signed cookies, CSRF, CI integrity pipeline, SRI.

Rules

• Checksums when receiving external components (go mod verify).
• encoding/gob for client data — no. JSON + signature or server-side storage.
• CI/CD: access control and mandatory review.
• CSRF: Go 1.25+ http.CrossOriginProtection (checks Origin/Referer, no tokens needed) or gorilla/csrf.
• Everything stored client-side is signed.

Recommended libraries & tools

• crypto/hmac & crypto/sha256 — data signing (stdlib).
• gorilla/securecookie — signed & encrypted cookies. One call to Encode/Decode.
• gorilla/csrf — CSRF protection via signed tokens. Middleware.
• Go 1.25+ http.CrossOriginProtection — built-in CSRF protection in stdlib.
• go mod verify — checks module integrity.

---

10. Logging & Alerting — record what helps investigation

OWASP: A09:2021 → A09:2025 — Security Logging & Alerting Failures. "Alerting" explicitly added to the name in 2025 — emphasis shifted to alerting, not just log collection.

Code examples: references/integrity-and-logging.md — log/slog setup, sensitive data avoidance (passwords, tokens, card numbers, reset tokens), audit middleware, Prometheus metrics.

Rules

• Log significant events: login/logout, failed attempts, permission changes, critical operations.
• Sensitive data never in logs. Passwords, tokens, card numbers.
• Structured logging: data as fields, not string concatenation. JSONHandler escapes values — injection impossible.
• Format suitable for automated analysis (JSON).
• Alert on anomalies: ">50 auth errors in 5 minutes", "spike in 5xx".

Recommended libraries & tools

• log/slog (Go 1.21+) — structured logging in stdlib. Preferred since 2024.
• zerolog — zero-allocation, JSON. For high-throughput (>100k req/s).
• zap — fast structured logger from Uber.
• sloggin / slog-echo — slog integration with frameworks.
• Prometheus + client_golang — metrics and alerting.
• Alerting pipeline — JSON logs → Vector or Promtail → Loki/Elasticsearch → Grafana.

---

11. External Objects — foreign data must not control logic

OWASP: cross-cutting — Insecure Design (A04/A06) + Injection (A03/A05).

Code examples: references/external-safety.md — path traversal, mass assignment, open redirect, external data validation, config whitelist.

Rules

• File paths: filepath.Abs + prefix check. Never filepath.Join with raw input alone. Go 1.24+ prefer os.Root/os.OpenRoot.
• Data updates: typed structs with field whitelist, not map[string]interface{}. Struct-based binding (Gin/Echo/Fiber) is whitelist by default.
• Redirect URL: relative paths only or domain whitelist.
• External API/webhook data: full validation before use.
• Configuration: whitelist of allowed keys.

Recommended libraries & tools

• path/filepath — filepath.Clean, filepath.Abs. Always verify with strings.HasPrefix.
• os.Root / os.OpenRoot (Go 1.24+) — prevents escaping root directory, even through symlinks.
• net/url — URL parsing and validation.
• go-playground/validator — dozens of built-in rules.
• Struct-based binding (Gin/Echo/Fiber) — JSON→struct ignores undeclared fields.

---

12. External Requests — limit server-side request destinations

OWASP: A10:2021 (SSRF) → consolidated into A01:2025.

Code examples: references/external-safety.md — URL/host whitelist, private IP blocking, safe HTTP client, DNS rebinding protection, response proxying.

Rules

• Server-side request URLs: whitelist of protocols, hosts, ports.
• Block requests to internal addresses (localhost, 169.254.x.x, private subnets). Check ALL resolved IPs.
• Never return raw external service responses to clients. Extract only needed data, limit size with io.LimitReader.
• If business logic allows, don't let users specify arbitrary URLs. Offer a choice.
• HTTP client: always with timeout, always with redirect validation, always check redirect targets for private IPs.

Recommended libraries & tools

• net — IP checking (net.ParseIP, net.ParseCIDR, net.IP.IsPrivate since Go 1.17).
• net/http — http.Client with Timeout and CheckRedirect. Default client has no timeout — always create your own.
• io.LimitReader — limit response body size. Without it, an attacker forces the server to download a gigabyte response.
• DNS rebinding protection — resolve once, check all IPs, connect via verified IP with explicit ServerName in tls.Config. For production, implement a retry loop over all resolved addresses — the first address may be unreachable.

---

---

13. AI-Assisted Development — process and code in the age of agents

OWASP: intersects with OWASP Top 10 for LLM Applications 2025 (LLM01 Prompt Injection, LLM02 Insecure Output Handling, LLM05 Improper Output Handling, LLM08 Excessive Agency) and all twelve sections above — because agent-written code falls into the same vulnerability categories as human-written code.

As of mid-2026, "ask an agent to write a handler" is as routine as searching Stack Overflow was a decade ago. One thing changed: the agent edits files, runs commands, reads dependencies and issues, sometimes deploys. This introduces two new risk classes:

1. Generated code quality. Models are trained on public repos — full of fmt.Sprintf in SQL, math/rand for tokens, text/template for HTML, and middleware with swallowed errors.
2. Process security. The agent reads external data (issue comments, READMEs of dependencies, web pages, MCP server responses) — any of these strings can contain instructions the model treats as commands (prompt injection, including indirect injection through repo file content).

Rules

• Agent context is pinned in the repository (AGENTS.md/SECURITY.md/skills), documenting security invariants.
• Specifications and tests are written before code generation; the agent implements them.
• Agent tools run in a sandbox without access to production secrets; tool permissions are granted minimally.
• AI-PRs pass the same CI gates as human ones: go vet, golangci-lint, govulncheck, go test -race. No exemptions for generated code.
• Sensitive changes (auth, crypto, data access, external requests) require mandatory human review.
• Agent tool-calls are logged at production audit-log level: tool, args, result, initiator.

Recommended libraries & tools

• AGENTS.md / SECURITY.md in the repo root — minimal but measurably effective way to set agent context.
• OWASP Top 10 for LLM Applications 2025 (genai.owasp.org) — separate risk list for LLM systems: Prompt Injection, Insecure Output Handling, Sensitive Information Disclosure, Excessive Agency.
• Sandboxing for the agent: dev containers, Docker-based runners, gVisor/Firecracker for strict isolation. Minimum: docker run --network=none --read-only for test execution.
• Tool-call logging: same slog JSON logs as in production, plus full prompt-input/output storage in a protected store — for post-incident prompt injection analysis.
• govulncheck and gosec on every AI-PR, mandatory (see sections 6 and Linters).
• Secret scanners in pre-commit: gitleaks, trufflehog. If the agent accidentally placed .env in a diff, the gate must be before push, not after.
• Prompt injection scanner for code and skills: ipi-check — scan everything fed to a coding agent.

---

Linters — automate what you shouldn't have to remember

go vet is necessary hygiene but not a full security suite. golangci-lint unifies 150+ linters.

Full configuration and CI setup: references/linters.md — .golangci.yml, linter reference table, gosec rules, CI/CD workflow, adoption strategy.

Key security linters

Linter	Primary concern
gosec	SQL injection, hardcoded secrets, weak crypto, path traversal, file permissions
bodyclose	Unclosed HTTP response bodies — TCP connection leaks
noctx	HTTP requests without context — goroutine leaks
sqlclosecheck + rowserrcheck	Unclosed DB resources + unchecked iteration errors
contextcheck	Broken context propagation chains
exhaustive	Missing enum values in switch
makezero	Slice initialization + append bugs

Running

go install github.com/golangci/golangci-lint/cmd/golangci-lint@latest
golangci-lint run --new-from-rev=HEAD~1   # changed files only
golangci-lint run ./...                   # full check (CI)

Adoption strategy

1. Start with golangci-lint run --new-from-rev=main — checks only new code.
2. Add linters one at a time, starting with gosec and errcheck.
3. Fix warnings in the current PR, don't accumulate technical debt.
4. If noisy on legitimate code, use //nolint:lintername // reason — don't disable globally.

---

Checklist

The full 60+ item checklist is in references/checklist.md, organized by OWASP category.