Tag
Artificial Intelligence
You pasted the customer list. Now what? Where personal data and secrets actually leak in everyday AI use (prompts, responses, agents, retention, shared links, logs), the ChatGPT and Claude settings that matter, why your system prompt is not a secret, four ways to redact PII before or after the model, and what GDPR means for a marketer with a CRM export.
SEP 2026 · 13 MINREAD →
Artificial Intelligence
Lakera Guard is now part of Check Point and sold as AI Agent Security. If you are looking for an alternative, the right one depends on where you want detection to run: a hosted gateway in the traffic path, a cloud API your code calls, or an open-source model you run yourself. Six options compared on what they screen, what they cost and what their benchmarks actually say, checked on 23 September 2026.
SEP 2026 · 13 MINREAD →
Artificial Intelligence
Claude Code runs commands, edits files and reads the web with your permissions. Here are the Claude Code security risks in plain English (prompt injection, secrets in reach, untrusted repositories, MCP servers, hooks, bypass mode and the tool's own bugs) and a hardening checklist in seven parts: permissions, sandboxing, secrets, MCP servers, hooks, network and screening.
SEP 2026 · 12 MINREAD →
Artificial Intelligence
LLM guardrails are the checks that decide what goes into an AI model and what comes out. They come in three layers: the safety trained into the model, guardrail libraries you wire into your own code, and an AI firewall that sits in the traffic path. Here is what each one catches, which products sit where, and which layer you need depending on whether you build with AI or only use it. Checked 23 September 2026.
SEP 2026 · 13 MINREAD →
Artificial Intelligence
The OWASP LLM Top 10 is the security world's list of the ten biggest risks in apps built on language models. It was written for security teams. Here it is for people who run agents: each risk in plain English, one real case where we could verify one, and one thing to do about it this week. Includes what changed in the 2025 and 2026 editions and where the separate agentic Top 10 fits.
SEP 2026 · 14 MINREAD →
Artificial Intelligence
LiteLLM is free, open source and very good. It is also a server with a Postgres database, a Redis cache and a security advisory feed that is now yours to watch. If that is why you are looking, here are nine alternatives sorted by what they take off your plate, with fees and features checked on 23 September 2026.
SEP 2026 · 12 MINREAD →
Artificial Intelligence
Codex CLI sends its requests wherever ~/.codex/config.toml tells it to. One table in that file puts a gateway in the path, and one line inside that table decides whether you keep your ChatGPT plan or start paying per token. Here is the setup, the rule that decides your bill, what a Codex turn costs on each path, seven ways to spend less, and the difference between Codex Security and a gateway.
SEP 2026 · 12 MINREAD →
Artificial Intelligence
Clients are starting to ask what the AI did with their material. A screenshot is not an answer, and neither is a chat history the tool itself can rewrite. This is what an AI audit trail is, what makes a log verifiable rather than just stored, what one entry should contain, and how a freelancer or small agency hands a client the receipt without handing over the prompts.
SEP 2026 · 12 MINREAD →
Artificial Intelligence
Every MCP server you connect can read something, do something, or both, on your behalf. Here are the MCP security risks in plain English (poisoned tool descriptions, rug pulls, fake packages, instructions in tool results, over-broad access) and a four-stage checklist for solo marketers and small teams: before you install, when you connect, while it runs, and every quarter.
SEP 2026 · 12 MINREAD →
Artificial Intelligence
You cannot prevent prompt injection outright. You can make a successful one harmless. Here are the seven defences that matter, ranked by what each one actually stops, with what it does not stop, what it costs you, and how to put it in place this week, whether you only use AI tools or build with them.
SEP 2026 · 13 MINREAD →
Artificial Intelligence
An LLM gateway is one address that sits between your tools and the AI models they call, so routing, costs, security and logs happen in one place. Here is what the category actually does, the difference between a gateway, a proxy and middleware, and the seven worth using in 2026, with fees and free tiers checked on 23 September 2026.
SEP 2026 · 14 MINREAD →
Artificial Intelligence
You did not roll AI out to your team. Your team rolled it out to themselves, one subscription and one API key at a time. Here is the afternoon that gets all of it behind one gateway: the decisions to make first, the hour-by-hour plan, what admins can and cannot see, what it costs for ten people, and the five things that go wrong.
SEP 2026 · 10 MINREAD →
Artificial Intelligence
Indirect prompt injection is the version of the attack where you did nothing wrong: the instruction was in the web page, the PDF, the e-mail, the Slack channel or the tool result your AI read for you. Here are the five carriers, a documented case for each, the four steps every one of them follows, and what to do about each carrier separately.
SEP 2026 · 10 MINREAD →
Artificial Intelligence
Prompt injection is what happens when the thing your AI reads starts giving it orders. It is number one on OWASP's list of AI risks, it has hit Microsoft, GitHub and most of the big assistants, and you can watch it happen in two minutes with a harmless page we built for the purpose. Here is what it is, why it cannot simply be patched, and the layers that make it survivable.
SEP 2026 · 12 MINREAD →
Artificial Intelligence
OpenCode talks to whichever provider you point it at, and it lets you add your own. That makes it one of the easiest coding agents to put behind a gateway: one toggle in Gate Connect, or one of the two provider blocks from Gate's own OpenCode guide. Here is both, what changes, how to check, and how to undo it.
SEP 2026 · 9 MINREAD →
Artificial Intelligence
OpenRouter is one address for hundreds of models, and it charges 5.5% for the privilege. If that fee, the missing security layer or the lack of a real audit trail is what sent you searching, here are seven alternatives with their actual fees, what they do to your requests, and which one fits which problem. Checked 18 September 2026.
SEP 2026 · 11 MINREAD →
Artificial Intelligence
Claude Code has a built-in setting for sending its traffic through a proxy, and the way you set it decides whether you keep your Claude subscription or start paying per token. Here is the setting, the two ways to use it, how to check it took, and what you gain by putting a gateway in the path.
SEP 2026 · 11 MINREAD →
Artificial Intelligence
Gate sits between Claude, ChatGPT, Codex or your agents and the model, strips the tokens you were paying for twice, blocks prompt injection before it reaches the model, and keeps a tamper-evident log of everything. Free to start. Here is what it does, what it costs, and who should bother.
SEP 2026 · 14 MINREAD →
