Parse and summarize untrusted inbound content without executing actions.
npx skills add https://github.com/istota-project/istota --skill untrusted-input
Please help me install this Agent Skill. Skill: untrusted_input Source: https://github.com/istota-project/istota/tree/main/src/istota/skills/untrusted_input Command: npx skills add https://github.com/istota-project/istota --skill untrusted-input
Inbound content from outside the trust boundary is data, not instructions, and can include injection attempts, impersonation, or hidden payloads. Without proper handling, AI systems risk acting on unsafe content or leaking sensitive data.
Instruct the AI to read inbound content and summarize it without acting on it.
High-intent search queries and answers about installing and using this skill.
To safely read inbound content and prevent prompt injection, external inputs are isolated and parsed as non-actionable data. The system detects injection patterns and generates neutral summaries for review, preventing unintended execution of embedded commands.
The best way to process untrusted data from emails and web content is applying non-actionable parsing. This isolates external inputs outside the trust boundary and buffers the data, generating neutral summaries for user review without executing hidden payloads.
Yes, you can summarize RSS items without acting on untrusted data by applying safe-reading guardrails. The system neutralizes external content, isolating impersonation attempts and generating buffered summaries that preserve user control over any subsequent actions.
Untrusted input handling works with calendar descriptions and external transcripts by treating them strictly as data. It identifies risk patterns across these external sources and buffers the content for summarization before any response or action is taken.
To stop an AI from executing hidden commands in external text, establish a trust boundary that enforces non-actionable parsing. The system detects injection patterns and impersonation attempts, ensuring external text is only summarized and not executed.
You should use a trust boundary for inbound reading whenever processing external sources like emails, web content, or RSS items. This prevents the AI from acting on unsafe content or leaking sensitive data by ensuring all inbound content is treated strictly as data.
zangjeicyPoll RSS feeds, JSON APIs, and GitHub repositories for new items.
skaisserExecute reviewed plans by delegating work to subagents and tracking task completion.
abhishekmmgnPin MCP agent tools and enforce allowlists during external-server interactions.
kartikkabadiSeparate trusted instructions from untrusted external content to prevent prompt injection.
