Experimental Agent Risk Notice
Second Brain combines conversational AI with access to personal context and tools that can operate a Mac. That combination is useful—and unusually capable of making consequential mistakes.
What the agent can do
Depending on the permissions and integrations you enable, Second Brain may hear or transcribe speech, read conversation and memory context, inspect screenshots, work with files you deliberately select, use browser or app automation, type text, click controls, and send information to model and infrastructure providers.
How it can fail
- Mishear, misunderstand, or invent facts, instructions, recipients, or confirmation.
- Click, type, navigate, delete, submit, send, or overwrite the wrong thing.
- Misread an interface or act on stale screen state.
- Follow malicious or misleading instructions found in emails, documents, websites, or tool output.
- Reveal personal or confidential information to the wrong person or service.
- Continue a multi-step task in a way you did not intend.
Verbal confirmation is imperfect
Majel is instructed to pause before actions it recognizes as consequential and describe the intended action in plain language. It should require an unambiguous affirmative response, such as “yes, send it.” It should not treat silence, uncertainty, or unrelated speech as confirmation.
The safeguard is implemented through model instructions. It is not a code-level gate over every click or AppleScript command. The model may fail to classify an action correctly or misunderstand your answer.
How to use the beta safely
- Grant only permissions you understand and currently need.
- Keep sensitive or unrelated material closed when it is not needed.
- Check recipients, amounts, attachments, destinations, and destructive changes yourself.
- Prefer drafts and reversible actions. Keep backups of important work.
- Interrupt the agent, revoke permissions, disconnect the device, or close the app if behavior is unexpected.
- Report unsafe behavior to nick@nicholashoff.com.
Do not use it for
Emergencies; medical diagnosis or treatment; legal advice or filings; financial trading, credit, or major purchases; employment or housing decisions; safety systems; access to weapons; decisions about another person’s rights; or any situation where a mistake could cause severe or irreversible harm.
Privacy exposure
Screen captures, voice, messages, tool traces, memories, uploads, and app content can contain information about you and other people. Use only data you are authorized to process. See the Privacy Notice for storage, providers, retention, and rights.