Every AI Agent Now Has Its Own Computer. Who Guards It?
Meta, xAI, OpenAI and Anthropic gave their agents their own computers. What they published to guard them, and the kernel layer below for agents you run.
An AI agent that runs commands, reads files, and calls tools is a new kind of user on your machine, one that can be talked into things by a poisoned file or an injected prompt. Prompt filters guard what it says. The risk is what it does.
InnerWarden puts the guardrail at the action layer. It screens every command the agent proposes, normalizes shell rewrites before matching, and on Linux, with Pro or Enterprise, the kernel Execution Gate refuses any program not on the agent's allowlist, scoped to the agent. Monitor first, enforce when you are ready. These posts cover the attacks, the MCP surface, and where enforcement has to live.
Meta, xAI, OpenAI and Anthropic gave their agents their own computers. What they published to guard them, and the kernel layer below for agents you run.
AI agent hardening in four layers: the agent's settings, containment, a command and tool-call guard, and kernel enforcement on Linux. Who provides each.
Claude Code hooks for security: how a PreToolUse hook blocks a tool call, two working examples, what a hook cannot stop, and what the Linux kernel adds.
Attackers defeat text command filters by rewriting a command into a form that means the same thing but does not match the string. The fix is two layers: normalize before matching, and read the real argv in-kernel at execve.
Most tools claim OWASP Agentic Top 10 coverage from the prompt layer. InnerWarden maps to all ten at the action layer with a reason chain, redaction, and a circuit breaker, and is honest about what is proven versus designed.
A poisoned input rewrites the agent's intent, and that intent lands on the host as execve, openat, or connect. Prompt-layer defenses live inside the context that just got poisoned. The damage becomes real at the syscall.
A model's refusal is a prediction an attacker can steer. The operating system's refusal is not a decision at all. Where enforcement has to live when agents touch real machines.
Auto mode is the right way to use a coding agent, and it executes commands with your permissions and no checkpoint. How to keep the speed and put a floor under it.
Tool poisoning, injected tool output, and command injection through tools: how each MCP attack works, why prompt defenses guard the wrong layer, and how to inspect the path itself.
A practical guide: put the free InnerWarden guard in front of your agent's commands with innerwarden serve, read the security context (Pro and Enterprise), and handle verdicts in Python and TypeScript.
Agentic Runtime Security is consolidating in 2026 around prompt-side defences. The load-bearing layer is the one downstream: what the agent does, not what it says.
How InnerWarden guards AI agents: MCP inspection with 85 ATR rules and taint tracking, proxy modes from advisory to kill, and alerts.
A step-by-step attack chain against an AI agent: prompt injection, tool poisoning, payload download, credential theft, and where each stage is caught or missed.
InnerWarden's AI isolation model: the model reads data and returns JSON recommendations, Rust validates and executes. The model never sees a shell.
GPU servers exposed to public APIs. Model-weight exfil, scraper bots, prompt-injection, supply-chain. eBPF + agent-guard MCP inspection + autonomous response.
How InnerWarden screens an OpenClaw agent's commands and tool calls, which part holds if the agent is turned, and how OpenClaw helps keep InnerWarden healthy.