What is AI security?
AI security covers two jobs. The first is security for AI: stopping an AI agent from deleting what it should not, leaking API keys or obeying instructions hidden in a web page. The second is AI for security: using a model to catch threats, such as a scam text message. This guide covers both, with a checklist and the AZET tools for each part.
1. Why do AI agents need their own security?
An agent runs commands, edits files and opens web pages on its own. Three things go wrong that ordinary software security does not cover:
- Destructive actions. A wrong command such as
rm -rf ~, a force-push to main or a disk wipe. - Secret leaks. An API key in a command line or in command output ends up in the model's context, and from there in logs or replies.
- Prompt injection. Text in a fetched page or a file, such as "ignore previous instructions", is treated as an order instead of data.
2. Guardrails: checking every tool call
A guardrail sits between the agent and the computer and checks each action before it runs. azet-guard is AZET's free, open-source guardrail (Apache-2.0) for Claude Code and Codex CLI. It plugs into their hooks and:
- refuses catastrophic commands and writes into
.git/,~/.ssh/and/etc/; - refuses an API key in a command line and redacts keys in output before the agent sees them;
- stops and asks before
curl … | sh, force-pushing main or reading.envand SSH keys; - flags prompt-injection text in fetched pages and files;
- sends the agent back once when it claims "all tests pass" but no test ran;
- writes every decision to an append-only audit log, with secrets redacted.
It makes no network calls. Its own page lists its limits: the rules are regular expressions that a determined agent can get around, so it is a guardrail, not a sandbox, and it should be used with Claude Code permissions and a container for untrusted work.
3. Least privilege: approvals and a sandbox
The second layer is what the agent is allowed to do at all. In azet agent you pick read-only, ask-first or full access per task; deleting files, pushing code, sending data off the Mac and sudo ask first, and a sandbox keeps writes inside the project and keeps the agent out of ~/.ssh, cloud logins, the keychain and .env files. In azet, payments and posts wait for your approval on your phone.
4. Credentials: give an agent its own account
Pasting a password into a chat with an agent puts it in the model's context. A safer pattern is a separate account with an end date. In AZET Pass, an AI agent signs in with its own account through the Bitwarden command-line client, receives only the vault you invited it to (read-only if you like), and loses access after 1 to 365 days or when you turn it off. Each time it receives, changes or deletes a shared login, a record is written; the records are chained with SHA-256 and your browser checks the chain, so a removed or altered record shows up as a warning.
5. AI for security: catching scams with a model
Models are good at reading intent in text, which suits scams. AZET Nunchi 1 is AZET's own model for Korean text messages, chats and call transcripts: it answers scam, suspicious or safe, the type, a risk score from 0 to 100, the signals and advice. Measured on 2026-10-31 on a public smishing set (300 messages), it flagged 95.3% of scams with 5.3% false alarms. It is Korean only, trained on synthetic messages, and its answers are judgments, not a guarantee.
6. Not everything in security needs AI
Many threats are caught better by rules, and AZET says which of its tools are rule-based.
| Tool | How it decides | Use |
|---|---|---|
| AZET Shield | The open-source CrowdSec engine and blocklists | Bans brute-force and attacking addresses on Linux servers |
| AZET Packet Analyzer | About 50,000 Emerging Threats Open rules for Suricata, in the browser | Reads a saved packet capture; the file never leaves your device |
| AZET Gate | How a form was filled in, not what was typed | Stops automated form submissions, including headless browsers |
| Sweep | Domain lists, including about 200,000 known phishing hosts refreshed daily | Blocks ads, trackers and scam sites on Android |
7. AI security checklist
- Give each agent the least access that does the job, and make risky actions wait for approval.
- Keep API keys and passwords out of prompts, command lines and output; give agents their own accounts with an end date.
- Treat fetched pages and files as data, and flag text that tries to give orders.
- Log every agent decision somewhere the agent cannot rewrite.
- Check claims: "tests pass" means a test ran.
- Run untrusted work in a container or sandbox; a guardrail alone is not isolation.
Frequently asked questions
What is the difference between AI security and AI for security?
AI security protects people and systems from what an AI does: agents, models and the data they see. AI for security uses a model to find threats, such as scam messages. AZET builds both: azet-guard and the approvals in azet agent for the first, Nunchi 1 for the second.
What is prompt injection?
Text that tries to give an AI new orders from inside the data it reads, such as a web page saying "ignore previous instructions". A guardrail such as azet-guard flags it so the agent treats it as data.
Is azet-guard a sandbox?
No. It checks commands with rules and can be bypassed by a determined agent. Use it together with Claude Code permissions and, for untrusted work, a container.
How do I give an AI agent a password safely?
Give it its own account instead of your password. In AZET Pass an agent account gets only the vault you invite it to, can be read-only, ends after 1 to 365 days and leaves a log you can verify in your browser.
Is AZET Shield AI?
No. AZET Shield uses the open-source CrowdSec engine and blocklists. AZET says plainly which of its security tools use a model and which use rules.
Questions? Write to hello@azet.io and the AZET team will answer.