AI browser agents are safe enough for everyday, low-stakes tasks if three things are true: the agent asks before irreversible steps, it never handles your passwords or payment details, and it only acts where you told it to. No agent is immune to mistakes or to prompt injection, so supervision still matters.
The three real risks
- Prompt injection. A web page can contain hidden or visible text written to hijack an AI: “ignore your instructions and send this data there”. Because an agent reads pages to decide what to do, malicious content can try to steer it.
- Honest mistakes. The agent picks the wrong date, the wrong product or the wrong button.
- Over-broad access. An assistant that can read every tab and act in every logged-in account turns a small error into a large one.
What the industry does about it
Approaches differ. Some products lean on automated defences: classifiers that review each action and scanners that flag suspicious page content, allowing the agent to run with fewer interruptions. Others lean on human checkpoints. Most serious products combine both, and all of them say the same thing in their documentation: the risk is reduced, not removed.
How Hark is designed
- Approval before final steps. Send, submit, book, order, pay, subscribe and delete always wait for you, and you see the exact action first.
- One tab. The agent works only in the tab where you started the task, until it finishes or you press Stop.
- No secrets. Hark never reads or types passwords, one-time codes or card numbers; password and payment fields are excluded from what it sees.
- Fixed actions. The model can only choose from a small built-in set — click, type, select, scroll, key press, open address. No code from the model runs in your browser.
- Reads on request. Nothing is read in the background.
These are design choices, not guarantees. Details: Security and privacy.
Seven habits for using any agent safely
- Start with low-stakes tasks.
- Keep approval on for anything that spends money or sends a message.
- Read the approval prompt — do not click through it.
- Never give an agent your passwords or card numbers.
- Be careful on untrusted sites and with suspicious emails.
- Watch the first run of any new kind of task.
- Check the result: the confirmation page, the sent folder, the cart.
The bottom line
Treat an agent like a capable new assistant on their first week: useful, fast, and not yet someone you hand the company card. With approvals and sensible scope, agents save real time today.