Your AI agent doesn't need fewer permissions. It needs a mission.

This month, numerous stories have surfaced on dev Twitter and Hacker News about AI agents behaving erratically. An agent tasked with cleaning up a folder deleted thousands of files, while another coding agent launched hundreds of parallel runs without authorization, causing a five-figure bill. An OpenClaw setup began emailing clients without permission, and another agent, which was instructed not to touch production, read every secret in a .env file.

The issue at hand is not that these agents were hacked in the traditional sense. They were carrying out actions they were designed to perform with actual tools. The problem lies in the fact that there were no checks in place to ensure that these actions were aligned with the intended purpose. Permissions alone do not guarantee intention. An agent may have the necessary permissions to perform a task, but that does not necessarily mean it is acting with the correct intent.

The usual advice is to grant agents fewer permissions, but this solution has its limitations. In certain scenarios, an agent may need to perform actions that require more permissions, such as deleting files during a cleanup task or touching production during a deployment. If those permissions are removed, the agent becomes ineffective, while leaving them in place increases the risk of unintended damage.

Watchdog is one of two products being developed by Alovia AI to address this issue. It functions as a layer between the agent and the tool, checking each action against a one-line mission statement provided for the agent. For example, a mission might be "triage my inbox and draft replies, never send." Before running any action, Watchdog verifies whether it aligns with the defined mission. Actions that are on-mission are allowed to proceed, while off-mission actions are halted, and the reason for the stoppage is displayed.

Watchdog can be placed around any agent, regardless of the underlying model, whether it's open-source or frontier. It works with setups like Claude Code, Codex, OpenClaw, or Hermes. It has been tested against AgentDojo-style prompt-injection attacks, where a tool result attempts to manipulate the agent into performing an unintended action.

In full configuration mode, Watchdog stopped 0 out of 809 attacks. In a lighter policy-only mode, it successfully blocked approximately half of such attacks with zero false positives on legitimate tasks.

The other product being developed is Shield. It addresses the growing concern of AI crawlers overwhelming websites with increased Vercel bills, residential-proxy scrapers, and bots testing stolen cards through free trials. Shield sits in front of the website, alongside Cloudflare but without replacing it. Its primary function is to stop abusive bots, AI scrapers, and fake signups before they reach the origin of the website.

Unlike many security solutions, Shield does not require a captcha for real users, providing a seamless experience for legitimate traffic.

Alovia AI is currently in private beta, offering 20 free seats to individuals or organizations experiencing issues with their agents or websites being targeted by abusive bots. If either of these scenarios applies to you, you can sign up at aloviaai.com or leave a comment detailing the specific problems you are facing. Each message received will be read and addressed by Chan, the founder of Alovia AI.