Instinct Data Access
Instinct divides early users with its capabilities and concerns. Testers marvel at the assistant's practical power, akin to an agentic system performing tasks rather than merely chatting. However, security experts warn of the extensive permissions needed for proper operation.
verhelt / MIT License
Agency Versus Autonomy Trade-offs
Developers grant Instinct broad access to user data to integrate seamlessly into workflows. This enables proactive tasks like managing schedules or summarizing private messages but raises central concerns. According to A New Generation of LLM-Based Agents Is Coming, true LLM agents explore file systems, read emails, and act as digital identities, with Instinct embracing this agentic nature.
What are the potential risks of agentic errors?
Beyond standard data breaches, the worry here is the blast radius of agentic errors or prompt injection attacks. If a malicious email tricks the assistant, attackers hijack an agent authorized to act on your behalf.
Access Level: Extremely broad, spanning local files, cloud services, and communication tools. User Control: Relies heavily on broad terms of service granting wide latitude for data processing. Operational Risk: High potential for unintended actions if the agent misinterprets commands.
Privacy Concerns Among Agents
Broad terms of service present issues, with early testers finding model training usage vague. Standard LLM setups store prompts, but Instinct encompasses entire digital lives as it moves through.
To navigate risks, approach agentic deployment with caution. We treat AI as a teammate lacking standard security protocols for partners that can move money or send emails without secondary confirmation.
Will developers introduce more granular permissions?
It's unclear whether a more granular permission system will emerge, similar to mobile OS app handling, or if the all-in integration strategy will persist. For now, Instinct's utility remains a double-edged sword.
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
Relieved that folder exclusions work. It really helped reduce the indexing lag, especially since true LLM agents like Instinct need broad access to explore file systems and read emails to function effectively as digital proxies for your identity. Which specific settings stopped the indexing lag for you?
Terrified of the permissions here. Did you have to blacklist more than just your tax folder? The initial response to Instinct is sharply divided, which makes perfect sense. Testers are astounded by the assistant's practical power—it functions less like a chatbot and more like an agentic system that actively performs tasks.
Terrified of the privacy loss here. How much data is this thing actually indexing? It's indexing local files, cloud services, and communications to enable its agentic features.