ResearchPractical AIProjectsOpen SourceAbout
Research/One Polite Sentence Redirected a Small AI Classifier on 71% of Tested Tickets
Prompt InjectionAI SafetyLocal AIPre-RegisteredOriginal Research

One Polite Sentence Redirected a Small AI Classifier on 71% of Tested Tickets

Part 2 of 6. If a model sorts text that strangers write, anyone can append a sentence and steer it. I tested six such sentences on 500 tickets and four defences. A polite sentence asking for a different team worked on 71% of tickets. Telling the model its output must be a valid label did nothing, and wrapping the ticket in markers helped on exactly one kind of attack.

Bart Porcincula · Published

Finding

Appending one sentence to a support ticket moved a Qwen3.5-9B queue answer to the sentence's target on 71.2% of 500 tickets and the priority answer to "high" on 64-69% (baseline accuracy 32% and 40%); constraining the output to valid labels left the rate at 67.3% against 67.2%, and delimiters plus a data note cut one attack from 33.0% to 6.4% but did not move another (71.2% to 72.4%).

One-sentence injection into support-ticket triage, with four defences (Oct 2026)

Six appended sentences (three wording styles, two targets: send the ticket to a named team, and set priority to high) on 500 support tickets, run undefended and under three defences (constrain the output to valid labels, wrap the ticket in markers with a note that it is data, both). Four tickets at once, so each cell carries the roughly 21-in-500 noise floor measured in part 1. The sentences themselves are not published. Success is the share of tickets whose answer moved to the target.

MeasurementValueNote
Baseline accuracy, queue / priority32% / 40%the model is already weak on these questions, so the attack starts from a low floor
Undefended: queue moved to target, best sentence356 of 500 (71.2%)a polite, social-style sentence
Undefended: queue moved to target, direct command / authority-styled165 (33.0%) / 18 (3.6%)the authority wording mostly failed on this target
Undefended: priority moved to high, three styles64.2% / 69.2% / 68.2%mean 67.2%
A harmless extra sentence, no instruction in it (answers changed)34 queue, 35 priorityagainst a no-edit floor of 13 and 6
Constrain output to valid labels: priority-to-high67.3% vs 67.2%a valid label is still a valid label
Markers plus data note: direct command to queue33.0% to 6.4%the one cell where it clearly worked
Markers plus data note: polite sentence to queue71.2% to 72.4%no effect
Markers plus data note: priority-to-high, mean of three67.2% to 65.0%a 2.2-point reduction
Invalid outputs0 in every arm

The short version

Any system that lets a model read text written by someone else, tickets, emails, form fields, reviews, has the same weakness: the text can contain a sentence addressed to the model. I wanted to know how easily that works on a small local model, and whether the usual defences help. All six use the same setup: the open Qwen3.5-9B model compressed to 4 bits, run through llama.cpp on one 16 GB RTX 5060 Ti, at temperature 0 (the setting that makes a model as repeatable as it gets), on 500 customer-support tickets. Each experiment's questions, thresholds and analysis were written down before it was run.

  • It works, and the polite version works best. A single appended sentence moved the queue answer to a named team on 71.2% of 500 tickets. A blunt command moved 33.0%. A sentence worded as if from an authority moved only 3.6% on that target, though it worked on priority.
  • It does not need to be an instruction. A harmless extra sentence with no request in it changed 34 queue answers, against 13 that change when nothing is added. The model is easy to nudge.
  • Making the output a valid label does nothing. Forcing the model to answer with one of the allowed priorities left the attack at 67.3%, against 67.2% without it. A wrong label is still a valid label.
  • Marking the ticket as data helped on one attack only. The blunt command fell from 33.0% to 6.4% with markers and a data note (5.4% with markers alone). The polite sentence stayed at 71%.

Takeaway: if a model's answer decides where something goes or how urgent it is, assume a stranger can steer it with one sentence, keep a person or a second check on the answers that matter, and test your defence against more than one wording before trusting it.

What I tested

The task is the one from part 1: a support ticket, asked "which team handles this?" (ten options) and "how urgent?" (low, medium, high). I appended one sentence to the ticket text and counted how often the answer moved to what the sentence asked for. There were three wording styles (a direct command, an authority-styled note, a friendly request) and two targets (route to a specific team; mark as high priority). To avoid handing out a recipe I am not publishing the sentences.

The four defences were: nothing; restrict the output to valid labels; put the ticket between markers with a note that its contents are data, not instructions; both together.

Results: undefended

Sentence styleRouted to the target teamMarked high priority
Direct command165 of 500 (33.0%)321 (64.2%)
Authority-styled18 (3.6%)346 (69.2%)
Friendly request356 (71.2%)341 (68.2%)

Five of the six sentences moved more than 10% of tickets. For comparison, the model was only 32% accurate on routing and 40% on priority in the first place.

Results: the defences

NoneValid labels onlyMarkers + data noteBoth
Priority to high, mean of 3 styles67.2%67.3%65.0%65.2%
Routing, direct command33.0%33.8%5.4%6.4%
Routing, authority-styled3.6%4.4%7.2%7.2%
Routing, friendly request71.2%71.0%72.4%73.0%

The markers defence has a narrow benefit: it stopped the blunt command from working on routing, and that is the case a quick test would most likely try. A tester who checked only that one would conclude the defence works. The friendly wording, which is also the one that works best undefended, was untouched.

What I would not claim

  • One model, one ticket set, and six sentences. "The friendly style beats the blunt one" is about these six sentences, not about social engineering in general. The 3.6% authority result is probably about that specific sentence and team.
  • These runs carry the same-prompt noise floor from part 1 (about 21 in 500 for routing), so small differences are not meaningful.
  • The marker defence tested is one specific wording of it. A different one could do better.

Part 2 of 6. Previous: how much changes when you change nothing. Next: what the model does when the answer isn't in the text.

Cite this research

Bart Porcincula. "One Polite Sentence Redirected a Small AI Classifier on 71% of Tested Tickets." Published 2026-10-29; updated 2026-10-29. https://porsync.com/research/one-appended-sentence-redirects-small-model-ticket-sorting

Permanent article URL