Bd Security

better-data: Security-sensitive changes

For library maintainers touching anything in better-data's security perimeter — Secret, EncryptionEngine, #[Sensitive], #[Encrypted], RequestSource guards, password handling. Mistakes in this perimeter aren't bugs that show up in tests; they're regressions that ship plaintext to disk or leak credentials in logs.

Misconception this skill corrects

> "I'll add a debug-mode that logs the encrypted value's plaintext when developer mode is on — it's only for development."

Don't. The security-feature-with-bypass is the worst outcome — caller assumes redaction is universal, log infrastructure picks up the "debug" path in production by accident, plaintext lands in CloudWatch / Sentry / wp-debug.log forever. The discipline is no-bypass: Secret's __toString returns '***' ([src/Secret.php:84-87](Secret.php)), jsonSerialize returns '***' ([src/Secret.php:89-92](Secret.php)), __debugInfo (controls var_dump / print_r) returns ['value' => '***'] ([src/Secret.php:99-103](Secret.php)), __serialize THROWS SecretSerializationException ([src/Secret.php:105-109](Secret.php)).

The throwing __serialize is deliberate — a caller serialized a Secret has already made a security-relevant mistake. Relaxing it to redact instead would silently let the bug ship. The exception forces them to either ->reveal() explicitly (audit point) or rethink the flow.

Other AI-prone misconceptions:

• "I'll cache the encryption key in a static property to avoid re-reading the constant on every call." Wrong — EncryptionEngine deliberately re-reads on every call ([src/Encryption/EncryptionEngine.php:53-54](EncryptionEngine.php)) so key rotation via BETTER_DATA_ENCRYPTION_KEY_PREVIOUS actually works. A process-long cache defeats rotation.
• "== and === are fine for comparing two Secrets; the constant-time stuff is paranoia." Wrong — string compare is timing-dependent and leaks length / first-byte equality through repeated probing. Secret::equals uses hash_equals ([src/Secret.php:78-82](Secret.php)). Always.
• "If decryption fails, return null and the caller falls back to the default." Wrong — silent failure on decrypt is worse than an exception. A caller that gets null instead of the expected secret may treat it as "user never set a key" and proceed. EncryptionEngine::decrypt throws DecryptionFailedException ([src/Encryption/EncryptionEngine.php:109,130](EncryptionEngine.php)).

When to use this skill

Trigger when ANY of the following is true:

• The diff touches src/Secret.php, src/Encryption/EncryptionEngine.php, src/Attribute/Encrypted.php, src/Attribute/Sensitive.php.
• The diff touches MetaKeyRegistry::register, route-owned-field handling, or RequestSource guards.
• New code introduces a Secret typed property OR #[Encrypted] attribute.
• New code calls EncryptionEngine::encrypt / decrypt / looksEncrypted.
• New code introduces password handling (user_pass, hash storage, comparison).
• Reviewing a PR that adds a "debug mode" / "verbose log" / "for development" flag near sensitive material.

Workflow

1. Threat model first (in the PR description or a code comment)

Write down explicitly:

• What is this change preventing? (e.g. "API tokens stored as plaintext in wp_options.")
• What remains un-prevented? (e.g. "Plaintext key in PHP memory between hydration and use; an attacker with PHP memory access can still read it.")
• What's the threat model? (e.g. "Database compromise, log file exposure. NOT defending against in-memory attackers — that requires HSM-grade key management.")

The goal: a future contributor reads the comment and knows whether their proposed change preserves or breaks the model.

2. Loud over silent

Every degradation must throw. The catalog:

| Condition | Throws | |---|---| | Encryption key constant not defined | RuntimeException | | decrypt called on garbage | DecryptionFailedException | | decrypt called with tampered ciphertext (GCM auth tag mismatch) | DecryptionFailedException | | Secret __serialize invoked | SecretSerializationException | | MetaKeyRegistry::register collision (one key, two DTOs) | RuntimeException | | RequestSource::noCollision finds an unexpected route-owned field in the body | RequestParamCollisionException | | strict mode + unknown DTO field in incoming data | UnknownFieldException |

If you add a security feature, add a corresponding throw. Don't ship a flag that converts the throw into a warning.

3. Symmetric end-to-end

Every encrypt has a matching decrypt, every redact has a matching reveal. Lookup table:

| Write side | Read side | |---|---| | SinkProjection::prepareValue encrypts when #[Encrypted] | AttributeDrivenHydrator decrypts when #[Encrypted] | | Presenter::toArray redacts Secret / #[Sensitive] | Explicit $secret->reveal() inside compute() closure | | OptionSink::projectForStorage encrypts | OptionSource decrypts | | __serialize throws | __unserialize throws — symmetric blockage |

If you add an encrypt path, you MUST add the matching decrypt path in the same PR. The reverse is also true. Asymmetry is a security regression because it preserves DATA but loses CONFIDENTIALITY.

4. Never cache the raw key

The pattern at [src/Encryption/EncryptionEngine.php:53-54](EncryptionEngine.php) is:

private const CONST_PRIMARY = 'BETTER_DATA_ENCRYPTION_KEY';
private const CONST_PREVIOUS = 'BETTER_DATA_ENCRYPTION_KEY_PREVIOUS';

Each encrypt / decrypt call reads the constant fresh. NO static property holds the resolved key. Reason: rotation. The runbook is:

1. Define BETTER_DATA_ENCRYPTION_KEY with the new key.
2. Define BETTER_DATA_ENCRYPTION_KEY_PREVIOUS with the old key.
3. Wait for the rotation period (lazy migration: every read decrypts under the old key, every write re-encrypts under the new).
4. Eventually undefine BETTER_DATA_ENCRYPTION_KEY_PREVIOUS.

If you cached the key in a static, step 1 doesn't take effect until the PHP process restarts — which can be hours after the constant changes. Don't cache.

5. Constant-time comparison

Any string comparison involving secret material:

// WRONG — timing oracle
if ($candidate === $stored) { /* ... */ }

// RIGHT
if (\hash_equals($stored, $candidate)) { /* ... */ }

Secret::equals is the canonical example ([src/Secret.php:78-82](Secret.php)). For password verification, use password_verify (which is constant-time by design).

6. Unit-test contract for security changes

Three probes minimum:

public function test_tampering_throws(): void
{
    $ciphertext = EncryptionEngine::encrypt('hello');
    // Flip one byte after the bd:v1: prefix
    $tampered = substr($ciphertext, 0, 8) . 'X' . substr($ciphertext, 9);

    $this->expectException(DecryptionFailedException::class);
    EncryptionEngine::decrypt($tampered, 'field');
}

public function test_missing_key_throws(): void
{
    // ensure the constant is undefined for this test
    $this->expectException(RuntimeException::class);
    EncryptionEngine::encrypt('hello');
}

public function test_secret_does_not_leak_via_dump_or_json_or_serialize(): void
{
    $secret = new Secret('sk_live_supersecret');

    $this->assertSame('***', (string) $secret);
    $this->assertSame('"***"', \json_encode($secret));
    $this->assertStringNotContainsString('sk_live', \print_r($secret, true));
    $this->assertStringNotContainsString('sk_live', \var_export($secret, true));

    $this->expectException(SecretSerializationException::class);
    \serialize($secret);
}

Add a leak probe for any new property / class that wraps secret material. The probe should print_r / var_dump / json_encode / serialize and assert the raw value isn't present.

7. Stress scenario for the WP boundary

The companion plugin's stress suite covers live-WP behavior:

• DTO hydrated from wp_options round-trips: option row contains ciphertext, hydrated DTO contains a Secret, reveal returns the original.
• REST response for a DTO with Secret shows '***' (consumer never sees plaintext over the wire).
• wp_options row inspection (raw SELECT) returns ciphertext, not plaintext.

wp better-data stress --filter Secret

A FAIL finding here blocks the change. A NOTE is acceptable but documented.

8. Run the full check

vendor/bin/phpunit
vendor/bin/phpstan analyse --memory-limit=1G
vendor/bin/php-cs-fixer fix
wp better-data stress

Critical rules

• Loud over silent. Every security degradation throws. No "soft fail with warning", no "log and return null".
• Symmetric end-to-end. Encrypt requires decrypt. Redact requires explicit reveal. New write path requires new read path.
• Never cache the encryption key. EncryptionEngine re-reads the constant per call; key rotation depends on it.
• hash_equals for any comparison involving secret material. == and === are timing oracles.
• Secret::__serialize throws — do not relax to redact. The exception forces explicit ->reveal() (audit point) or rethink.
• No "debug mode" that logs plaintext. Even conditionally. The flag inevitably runs in production.
• Threat model in the PR description or a code comment. Future contributors need to know what's protected and what isn't.
• Tamper probe + leak probe in unit tests for any change touching Secret / EncryptionEngine.
• MetaKeyRegistry::register collisions throw. Two DTOs claiming the same meta_key is a programming error, not a coexistence-via-overwrite scenario.
• RequestSource strict mode whitelist throws on unknown fields. Don't relax to "ignore unknown" — the strict mode IS the security boundary.
• Constant-time comparison for password equality. Use password_verify for hashes; hash_equals for raw secret comparison.

Common mistakes

// WRONG — debug-mode plaintext logging
public function decrypt(string $ciphertext, string $field): string
{
    $plaintext = self::actualDecrypt($ciphertext);
    if (\defined('BETTER_DATA_DEBUG') && BETTER_DATA_DEBUG) {
        \error_log("Decrypted {$field}: {$plaintext}");  // 🔴 plaintext to log
    }
    return $plaintext;
}

// RIGHT — no plaintext-leaking branches at all

// WRONG — silent failure on decrypt
public function decrypt(string $ciphertext, string $field): ?string
{
    try {
        return self::actualDecrypt($ciphertext);
    } catch (\Throwable) {
        return null;  // 🔴 caller can't tell garbage data from missing data
    }
}

// RIGHT — let it throw
public function decrypt(string $ciphertext, string $field): string
{
    return self::actualDecrypt($ciphertext);
    // throws DecryptionFailedException on tampered / wrong-key / garbage
}

// WRONG — caching the key
class EncryptionEngine
{
    private static ?string $cachedKey = null;

    private static function getKey(): string
    {
        return self::$cachedKey ??= \constant(self::CONST_PRIMARY);  // 🔴 rotation broken
    }
}

// RIGHT — fresh read every call
private static function getKey(): string
{
    return \constant(self::CONST_PRIMARY);  // O(1) constant lookup; not worth caching
}

// WRONG — equality with ===
if ($candidate === $secret->reveal()) {  // 🔴 timing oracle
    grant_access();
}

// RIGHT
if ($secret->equals($candidate)) {  // hash_equals under the hood
    grant_access();
}

// WRONG — relaxing __serialize to redact
public function __serialize(): array
{
    return ['value' => '***'];  // 🔴 silent: caller's serialized blob looks "fine"
}

// RIGHT — throw, force explicit reveal at the call site
public function __serialize(): array
{
    throw SecretSerializationException::forSerialize();
}

// WRONG — encrypt-only
class SinkProjection {
    static function prepareValue(...): mixed {
        if ($encrypted) return EncryptionEngine::encrypt($value);
        return $value;
    }
}
class AttributeDrivenHydrator {
    static function hydrate(...): DataObject {
        // forgot to call EncryptionEngine::decrypt
        return $dto;
    }
}
// Result: stored ciphertext, hydrated ciphertext-as-string. Looks like garbage on read.

// RIGHT — symmetric pair lands in the same PR

Cross-references

• Run bd-attribute when adding a new security-relevant attribute (e.g. #[RedactInLogs]) — wire write-side and read-side together.
• Run bd-sink + bd-source-adapter when extending encryption to a new store — both sides land in the same PR.
• Run bd-better-route-bridge if the change affects how secret-bearing DTOs cross the REST boundary — verify Presenter::rest() redaction is still in effect.

What this skill does NOT cover

• Choosing a different cipher (aes-256-gcm is the current default; changing it is out of scope and would be a breaking change for stored ciphertexts).
• Key management infrastructure (HSM, AWS KMS, Vault). Library reads PHP constants; integrating with a key store is a consumer concern.
• HTTPS / TLS configuration. WP / hosting layer.
• Login auth / session handling. WP-level concerns; better-data consumes the resolved current user.
• Side-channel attacks beyond timing (cache, branch prediction, …). Not in threat model.
• GDPR / data-subject-rights tooling. The library is data-shape-only; export/erasure is a consumer responsibility.

References

• Secret: [libraries/better-data/src/Secret.php:60-120](Secret.php) — final class Secret, __toString returns '***' (line 84-87), jsonSerialize returns '***' (89-92), __debugInfo returns ['value' => '***'] (99-103), __serialize throws (105-109), __unserialize throws (111-117), equals uses hash_equals (78-82), reveal (70-73).
• EncryptionEngine: [libraries/better-data/src/Encryption/EncryptionEngine.php](EncryptionEngine.php) — BETTER_DATA_ENCRYPTION_KEY (line 53), BETTER_DATA_ENCRYPTION_KEY_PREVIOUS (54), looksEncrypted (137), decrypt throws DecryptionFailedException (109, 130).
• #[Encrypted] attribute: [libraries/better-data/src/Attribute/Encrypted.php](Encrypted.php) — pure data carrier; logic is in projection / hydration.
• #[Sensitive] attribute: [libraries/better-data/src/Attribute/Sensitive.php](Sensitive.php) — read by Presenter::sensitiveFieldNames.
• Idempotent decrypt: [libraries/better-data/src/DataObject.php:173-184](DataObject.php) — looksEncrypted + Encrypted attribute check.
• MetaKeyRegistry: [libraries/better-data/src/Registration/MetaKeyRegistry.php](MetaKeyRegistry.php) — collision throws.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Lonsdale201
• Source: Lonsdale201/wp-agent-skills
• License: MIT

Install and usage instructions live in the source repository linked above.