JevMade hello@JevMade.com
← Back to guides

JevMade field notes / Evaluation postmortem

Finding the safety limits of an AI email sorting tool

This guide explains how a developer tested an AI email manager. It shows how the software sorts messages, handles hidden tricks, and why some safety tests fail to prove the system actually stops threats.

Original by 0xShin0221Guardrails

Listen to this guide

JevMade’s plain-English explanation

0:00 /

AI narration

Credits

“openpoke-meets-jev” by 0xShin0221. Read the original source.

This expanded guide is an AI-narrated adaptation prepared by JevMade. It expands the source’s essential ideas, examples and caveats in JevMade’s own words and is not a word-for-word reading. The synthetic voice does not imitate the author or imply their endorsement.

Our summary

This project is an AI email manager that reads incoming messages and alerts you about important ones. Instead of asking a standard AI assistant to read every message, it uses a specialized tool called Jev to quickly sort emails by choosing from set options.

The software asks Jev to rate email importance and check for hidden instructions meant to trick the AI. If a message looks suspicious, the software quarantines it. In one safety test, attacks failed only because the main AI assistant simply ignored the tricks, not because the safety checks stopped them.

This guide is useful for people building AI email tools, but the author notes several missing measurements. The project does not yet know if the new sorting method is actually accurate or if it saves money overall. The safety checks were also not proven to stop attacks.

Key takeaways

  1. A specialized AI tool can sort emails by choosing options rather than writing out text.
  2. Automatically deleting suspicious emails can allow attackers to intentionally hide your real messages.
  3. A failed attack in a test does not prove your safety checks actually stopped the threat.

The author warns that the safety test results only show the AI assistant ignoring the attacks, not the safety checks working. The project also lacks a checked set of examples to prove the sorting tool is accurate.

Repository README · Source reviewed

Read the original guide Opens the author’s site in a new tab.