Practical security for AI that can act.
Detailed guides for people giving AI access to real files, accounts, repositories and infrastructure.
How to secure AI coding agents
A practical security model for Claude Code, Codex, Cursor, Devin and every other agent that can edit files, install packages, run commands or reach production systems.
Secure the authority around the agent, inspect everything it changes and keep narrow deterministic controls around actions that could cause irreversible damage.
Read 14 minAnswers you can use.
Threat models, checklists and operating guidance grounded in how current agent tools actually work.
How to use AI desktop apps safely with real files and accounts
Desktop AI can read folders, move documents, use connected services and act with your authority. These practical controls keep useful automation from becoming an expensive mistake.
Give AI enough access to complete the task, not standing authority over every file, account and connected service.
MCP security checklist: 12 checks before you connect a server
MCP turns useful context into executable authority. This checklist covers server trust, tool permissions, authentication, prompt injection, secrets and ongoing monitoring.
How to stop AI agents running destructive commands
Permission fatigue makes broad approval inevitable. Here is how to keep agents useful while protecting files, Git history, databases, containers and cloud infrastructure.
What is slopsquatting? How AI package hallucinations become attacks
A coding model invents a plausible package name. An attacker publishes it. A developer installs it. Learn how the attack works and how to stop it before installation.
Claude Code security: permissions, hooks and the gaps around them
Claude Code has strong native controls. Learn what its permission system, sandbox and hooks do well, when teams bypass them and where an independent layer still matters.
Claude Security, Codex Security and deterministic scanners
Frontier security agents and deterministic scanners solve different problems. This comparison shows where each is strongest and how to combine them without paying to scan everything twice.
The architecture behind the controls.
Shorter essays on deterministic security, agent authority and the software supply chain.
Permission prompts are not a security boundary
Agent permissions are useful, but a prompt that developers routinely bypass cannot carry the whole security model. High-impact actions need a narrower independent boundary.
Keep provider permissions. Add an independent policy layer that remains useful after a developer grants broad access.
Why AI development needs a deterministic security layer
Frontier models can reason about risks that rules have never seen. They are not a stable substitute for reproducible policy, continuous checks or audit evidence.
What changes when an AI agent gets your shell
The risk is no longer limited to a bad code suggestion. A capable agent can touch files, databases, cloud tooling, deployment systems and every credential available to the session.
AI agents change the software supply chain, not just the code
A dependency suggestion, MCP server, skill or persistent instruction can alter the trust boundary faster than a source-code edit. They deserve the same continuous scrutiny.
The FAQ covers deployment, provider compatibility, runtime limits and how CodeMarine works with frontier security tools.
Open the full FAQ →Build with the frontier. Keep an independent watch.
See how CodeMarine fits around the tools your team already uses.
