Security Engineering

All five mainstream reverse proxies will work. Choose on operating model, meaning where configuration comes from, not on benchmark numbers you will never hit.

An MCP server hands a model the power to act. Here is how to secure it: authN/authZ per invocation, least-privilege tools, untrusted-output handling, brokered secrets, and audit logs.

A first pass at LLM red-teaming: the four failure classes, a starter probe set, direct vs indirect injection, scoring, guardrail limits, and turning it into a CI regression suite.

An agent is a new principal. Do not give it the user's cookie or the server's root key. Per-action auth, brokered secrets, and an audit id.

Deploy TLS inspection only with managed devices, a written bypass list, and a defensible retention policy. Here is what it sees, what it breaks, and where it stops working.

Default-deny egress at a proxy the agent cannot route around is the highest-leverage single control for agent security. An allowlist alone will not stop exfiltration.

Gmail's blue checkmark requires BIMI, DMARC at enforcement, a registered trademark, and a VMC costing $750 to $1,400 a year. Full requirements, steps, costs, and when to skip it.

Reaching p=reject takes six to eight weeks and one hard prerequisite: naming every system that sends as your domain. The staged rollout, the alignment rule people miss, and what breaks.

Web Bot Auth signs agent requests with a verifiable key, turning "this claims to be ChatGPT" into proof. The drafts, the CDN path, origin verification, and the policy that has to come first.

Sigstore proves who built it, SLSA provenance proves how. Generation is nearly free on hosted CI; verification with a real policy is the project. Levels, steps, and the silent failures.

Signing what you produce, verifying what you ingest, and deciding what absence means. The pipeline audit that decides whether any of it survives to the browser.

A long-lived key in a repo or an agent env is a master key you cannot inventory. Workload identity, short-lived tokens, and brokered secrets are the replacement.

Gmail's checkmark proves a trademark. X's proves a payment cleared. LinkedIn's proves an ID matched a face. What each badge checks, what it costs to counterfeit, and why none of them make a message safe.