AI Concerns

Map/D3/3.1

I · D3 · 3.1

Indirect prompt injection

Malicious instructions hidden in documents, emails, web pages the model reads.

Major · 4Documented at scaleNowRisk 20

Severity

4/5

Likelihood

5/5

In the hierarchy

  1. Part I: Technical & System Concerns
  2. D3. Security & Adversarial Risk
  3. 3.1 Prompt injection & jailbreaks

Virtues and principles to explore

Held as inquiry, not as a verdict.

  • Stewardship

    Service

    Hold what is built in trust for those who will live with it.

  • Prudence

    Excellence

    Small, reversible steps before irreversible ones.

  • Accountability

    Trust

    Name who holds the consequence before it is needed.

Starter questions

Written to probe curiosity and learning, not accusation.

  1. 01What would Accountability require of us here, before we proceed?
  2. 02What might we be missing if we treat “Indirect prompt injection” only as a technical problem?
  3. 03If a colleague raised this concern tomorrow, how would we receive it as inquiry rather than accusation?
  4. 04What capacity would we need — in people, in the institution, in the tools — to meet this well?

Also on this branch

Frameworks and sources

Writing and incidents

Open indexes first. Then, if you wish, ask Grok to search the live web for this concern — one request, cached for the rest of this session.

NewsScholarIncidentsOECD monitorMIT repositoryWikipedia
See it on the mapTest an initiative against this

Attacks on and through AI systems: injection, jailbreaks, cyber offense.