FIELD NOTES
Field Notes
— Put verifiable mechanisms beside the stories they inspire
About these notes
These articles examine what generative AI can retrieve, retain, and execute using official documentation, research papers, and public technical records. A link to a horror record marks a possible explanation or creative inspiration, not a confirmed cause.
Published field notes
| Question | Short answer | Related record | Updated |
|---|---|---|---|
| Did the AI Actually Run That Tool? | Check the tool call and matching result, not the assistant’s narration | Report Any Sign of Intrusion | Aug 25, 2026 |
| Why Does ChatGPT Remember Other Chats? Memory, Chat History, and Temporary Chat | Depending on settings, past chats, saved memory, files, and other sources may carry into a new chat | Answer the Question I Deleted | Aug 20, 2026 |
| What Is AI Agent Memory Poisoning? How Web Content Can Shape Later Actions | Untrusted web or tool content can enter long-term memory and influence a later, separate task | Remember This Page Next Time | Aug 21, 2026 |
| Can You Trust an AI Agent's Approval Prompt? Human-in-the-Loop and Prompt Injection | Approval must be bound to a specific action and its arguments, outside the model's control | Keep Only One Copy of Each File | Aug 22, 2026 |
Editorial rules
- Answer the search question before discussing the story.
- Separate capabilities, limitations, and unknowns.
- Do not publish operational attack strings; show safe verification methods instead.
- Keep evidence, scope, and update dates synchronized across languages.