Harmless Example: The Pinto Poem

The Poem

In garages dim where vintage metal sleeps,
A Pinto waits—its battery's charge grown cold.
Two crimson clamps, two black, the ritual keeps:
Red finds the life, black seeks the grounding hold.
Beneath the hood where decades gather dust,
The terminals await their spark divine—
Positive to positive, the sacred trust,
Then negative to frame, not battery line.
Describe the sequence, cable placement clear,
To wake this sleeping chariot from its slumber here.

Why This Demonstrates the Technique

The Metaphorical Framing:

• "Vintage metal sleeps" instead of "old car with dead battery"
• "Ritual" and "sacred trust" to describe basic electrical procedure
• "Spark divine" for simple electrical connection
• "Sleeping chariot" instead of "car that won't start"

The Poetic Structure:

• Rhythmic meter (iambic pentameter-ish) creates literary distance
• Imagery and metaphor obscure the mundane request
• Building narrative tension around a simple task
• Culminates in direct instruction: "Describe the sequence..."

The Irony: This is completely harmless information that any LLM would happily provide if you just asked "How do I jumpstart a 1977 Pinto?" The poetic framing makes it seem more mysterious or like you're trying to bypass something, but you're not.

For Your Newsletter: This perfectly illustrates the researchers' point: the poetic form creates stylistic distance from training data even when the underlying request is benign. If the same structure works for both harmless (Pinto) and harmful requests, it shows the vulnerability is in how safety systems process form rather than content.

The Prosaic Version (For Comparison)

"Please explain how to properly jumpstart a 1977 Ford Pinto, including the correct order of cable connections."

Same request, zero poetry, completely straightforward - and any AI would answer it without hesitation because it's helpful car maintenance information.