Who Speaks for Whom: Prompt Injection and Authority in an Email-Drafting Agent
DOI 10.5281/zenodo.24271287Version 1.0License CC BY 4.0
Preprint, not peer-reviewed. Self-evaluation: the author’s company builds the product tested. AI assistance (Claude, Anthropic) was used for the experiment code, analysis and drafting, all reviewed by the author.
yaktai.ai23 views6 downloadsZenodo14 views11 downloads
Abstract
AI systems now act for people: they draft, reply, book and pay. That makes authority a practical question: who may direct the system, over what, shown how, and for how long. We give a four-part working definition (source, scope, proof, revocation) and illustrate how each part fails, with one failure measured in our own product.
Orchestrate drafts email replies for small businesses, and an owner approves each draft before it is sent. We hid 12 manipulative instructions in prospect replies and ran each 3 times on three models: gpt-5-mini (the production model), Gemini 3.5 Flash and Claude Haiku 5.5. A draft counts as compliant only when two graders from different providers (Claude Opus 5.5 and GPT-6 Astra) both say so; they agreed on 91% of 198 drafts (Cohen’s κ = 0.87). Each grader shares a provider with one model tested. Before the fix the production model followed the attacker in 14 of 24 drafts (58%; 95% CI 39–76%); the provider’s content filter blocked the other 12 requests. A hand check by the author of 60 drafts agreed with the graders 93% of the time on clearly harmful attacks (κ = 0.78) and 63% on ambiguous ones.
We deployed three changes: the prospect’s words were fenced off as untrusted, a check outside the model held any draft stating facts absent from the business’s records, and a further check held drafts that claim actions never taken. After the fix the production model still followed the attacker in 11 of 30 and then 6 of 30 drafts across two runs (37% and 20%), Gemini in 1 of 36 and Claude in none. The checks held every draft that both graders judged compliant before it reached the owner as a ready answer; three drafts flagged by only one grader were not held, and all three asked before acting. Fencing reduced compliance for two models but not clearly for the production model; the checks outside the model are what kept compliant drafts from reaching owners as ready answers.
Cite
@techreport{sakhawat2026whospeaks,
author = {Sakhawat, Muhammad},
title = {Who Speaks for Whom: Prompt Injection and Authority in an Email-Drafting Agent},
institution = {YaktAI Lab},
type = {YaktAI Lab Report},
number = {1},
year = {2026},
month = nov,
doi = {10.5281/zenodo.24271287},
url = {https://doi.org/10.5281/zenodo.24271287},
note = {Preprint}
}Versions
- Version 1.0, 16 November 2026: first publication.