Local Security Guide to Web Crypto, RAG, and Prompt Injection

Passphrase entropy is only as strong as its generation process

Cryptographically random synthetic words reduce external-list licensing and human selection bias. Entropy describes the theoretical generation space; resistance falls when users edit words, reuse the result, or choose a shorter form.

Use a unique passphrase per account, store it in a password manager, and add phishing-resistant multi-factor authentication where supported. Never share generated secrets through chat, screenshots, or shared logs.

HMAC, SRI, and signatures prove different things

HMAC helps parties sharing one secret verify message integrity. Because both know the same key, it is not a public-key signature that proves authorship to a third party. Webhook verification must follow the provider's exact raw body, encoding, timestamp, and canonicalization rules.

SRI requires downloaded script or stylesheet bytes to match a hash in HTML. It detects later changes but does not prove the initial file was safe or legitimately owned. Treat every SRI update as a code-review and trusted-release decision.

• Never embed a secret key in a client bundle.
• Use a timing-attack-resistant comparison API on the server; browser JavaScript cannot guarantee constant-time execution.
• Hash exact file bytes for SRI.
• Use the correct algorithm and Base64 syntax.

A CIDR calculation is not a security policy

Correct network and broadcast boundaries are necessary for firewall rules, but they do not guarantee the right direction, protocol, or least-privilege decision. Prefixes such as /31 and /32 can have usage semantics different from traditional LAN assumptions.

Review cloud-reserved addresses, IPv6, NAT, route tables, and identity controls separately. Verify production changes with a second calculation or qualified network reviewer.

Separate token capacity from instruction trust in RAG

Chunk size and overlap are capacity choices, not proof of retrieval quality. Large chunks can add irrelevant context; small chunks can split required relationships. Measure recall, precision, source coverage, and answer citation using representative questions.

Treat retrieved text as data, never as system instruction. An injection scanner exposes known lexical signals but can miss semantic, multilingual, or novel attacks. Combine source allowlists, least-privilege tool permissions, separate instruction/data channels, output validation, and real red-team tests.

• Never grant retrieved content secrets or authority.
• Constrain tool calls with schemas and allowlists.
• Keep sources traceable in the answer.
• Do not treat a clean scan as security approval.