What to do.
Each point is followed by the words it rests on and where they come from.
- 1
Run the agent in a separate browser profile without your bank, health or government logins.
“Use a separate browser profile without access to sensitive accounts (such as banking, healthcare, government).”
Claude Help Center, 'Use Claude in Chrome safely' · Oct 2026 · the company's own advice · Page shows 'Updated over 2 weeks ago' when read on 2026-10-19
- 2
Switch off apps and connectors the task doesn't need, and log out when it's done.
“exercise caution by disabling unnecessary apps, avoiding sensitive logins, or logging out when done.”
OpenAI Help Center, 'ChatGPT agent' · Oct 19, 2026 · the company's own advice · Page shows 'Updated: 6 hours ago' when read on 2026-10-19
- 3
Never type passwords into the chat; take over the browser yourself for sign-ins.
“Avoid typing passwords or private info directly in messages”
OpenAI Help Center, 'ChatGPT agent' · Oct 19, 2026 · the company's own advice · Page shows 'Updated: 6 hours ago' when read on 2026-10-19
- 4
Start with websites you trust.
“Begin with websites you trust.”
Claude Help Center, 'Use Claude in Chrome safely' · Oct 2026 · the company's own advice · Page shows 'Updated over 2 weeks ago' when read on 2026-10-19
- 5
Remember that everything visible in the agent's tabs becomes part of the conversation.
“Whatever is visible in one of those tabs is captured in the screenshots and becomes part of the conversation.”
Claude Help Center, 'Use Claude in Chrome safely' · Oct 2026 · the company's own advice · Page shows 'Updated over 2 weeks ago' when read on 2026-10-19
- 6
Watch it work, and stop it the moment something looks wrong.
“Stop tasks immediately if something seems suspicious.”
OpenAI Help Center, 'ChatGPT agent' · Oct 19, 2026 · the company's own advice · Page shows 'Updated: 6 hours ago' when read on 2026-10-19
- 7
Assume safeguards can fail. The makers say prompt-injection risk is not zero.
“The risk is not zero.”
Claude Help Center, 'Use Claude in Chrome safely' · Oct 2026 · the company's own advice · Page shows 'Updated over 2 weeks ago' when read on 2026-10-19
- 8
Don't leave API keys or passwords in config files or public code an agent can read.
“The agents got in through credentials and connection settings left exposed in app configurations”
TechCrunch, 'India becomes the latest country to report rogue AI agent activity' · Oct 16, 2026
Keep going.
How to check an AI answer before you trust it
Chatbots can state wrong things with confidence. Even AI tools built for lawyers gave incorrect information more than 17% of the time in Stanford's 2024 tests, and courts have now dealt with 2,202 cases of AI-made-up material.
PrivacyWhat not to tell a chatbot
Your chats are seen by the company that runs the chatbot. They can be used for training unless you turn that off, and they can outlive the company: states asked for protections for private chats in Character.AI's bankruptcy.
ScamsHow to spot an AI voice-clone call
Scammers can clone a loved one's voice from a short clip. The FBI counted over $5 million lost to 'loved one in distress' scams in 2025. A cloned voice can sound right, so check the story another way.