The Test Kitchen · Read further
Sources & how they were checked
Each entry says what the source supports on this site and how it was verified on 2026-10-26. Where a link was not opened in this review, it says so.
Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)
NIST · July 2024
Voluntary companion to the AI RMF naming twelve risks unique to or worsened by generative AI, including confabulation, data privacy and information security. Used here for the confabulation and leakage framing.
CheckPublication record located via search on 26 Oct 2026; details corroborated by secondary explainers. DOI: 10.6028/NIST.AI.600-1.
LLM01:2025 Prompt Injection
OWASP GenAI Security Project · Page modified 17 Apr 2025
Defines direct and indirect injection, lists mitigations (constrain behaviour, validate outputs, least privilege, human approval, segregate external content, adversarial testing) and says fool-proof prevention is unclear.
CheckOpened and read on 26 Oct 2026.
OWASP Top 10 for LLM Applications 2025
OWASP GenAI Security Project · 2025
The full 2025 list, including Sensitive Information Disclosure (LLM02), System Prompt Leakage (LLM07) and Misinformation (LLM09).
CheckIndex linked from the LLM01 page opened on 26 Oct 2026.
Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection
Greshake et al. · arXiv, Feb 2023
The paper that popularised indirect prompt injection through retrieved content. Cited in OWASP's LLM01 references.
CheckReference listed by OWASP; abstract page not separately opened.
Lost in the Middle: How Language Models Use Long Contexts
Liu et al. · arXiv, Jul 2023 (TACL 2024)
Found that models used information at the start and end of long contexts more reliably than in the middle. Motivates narrow questions and page references for long documents.
CheckWell-known reference; link not rechecked in this session.
The 2026 AI Index Report
Stanford HAI · Released 13 Apr 2026
Annual data report on AI capability, adoption, policy and responsible-AI practice. Background on why evaluation and measurement deserve attention; no figure from it is quoted on this site.
CheckReport page located via search on 26 Oct 2026; release date from coverage of the launch.
Prompt engineering overview (Claude documentation)
Anthropic · Living documentation
Vendor guidance on clarity, examples, XML-style tags and defining success criteria before iterating. One vendor's practice, not a universal standard.
CheckLink not opened in this session.
AML.T0051 LLM Prompt Injection
MITRE ATLAS · Living knowledge base
Adversary technique entry for direct prompt injection, useful for threat modelling.
CheckLinked from OWASP LLM01; not separately opened.
What this site does not claim
- No search-volume data; the keywords targeted are hypotheses.
- No statistics from the AI Index are quoted.
- No live model outputs; all examples are constructed.
- No validation that the eight stations predict output quality.
Related worlds
Shows the other half of this kitchen's argument: a browser tool that refuses to fake AI. Its 'Ask the archive' ranks by visible keywords and says so.Opened 26 Oct 2026; redirects to its canonical synthesis.sahiltayal.com.The Hearsay Department
A playable lesson in how a partial observation becomes a confident report. Confabulation is the machine version of hearsay.Opened 26 Oct 2026.The Archive of Attempts
Iteration done honestly: intention, procedure, observation and the conclusions that do not follow. The method behind 'test, then revise'.Opened 26 Oct 2026.The Sahilverse Atlas
The full catalogue. The Agency Machine (on intention vs action) is listed there; its site blocked automated checking, so it is not linked here as verified.Opened 26 Oct 2026.If AI Can Do My Thinking, Who Am I? (essay)
Why this kitchen keeps the human as the taster: authorship and judgment survive delegation only if you check the dish.Listed as a verified article on sahiltayal.com; not opened in this session.