Bluesky Thread

your embeddings are not safe!

View original thread
your embeddings are not safe!

every prompt directly maps to its embedding and back. they’re isomorphic

SipIt is a linear time algorithm for quickly and efficiently extracting input text from embeddings

www.arxiv.org/abs/2510.15511
Schematic showing how changes in prompts map to changes in a model’s latent space.

Left panel, titled “Prompt Space”: a beige, wavy sheet with thin contour lines and many black dots. Two dots are labeled x (lower middle, with an arrow pointing to it) and x′ (upper right). A short dotted segment between them is labeled δ.

Right panel, titled “Latent Space”: a square grid with x–y axes (a right-angle mark at the origin) and the label ℝᵈ. Two points are shown: z (lower) and z′ (upper right). A short dotted segment between them is labeled ε.

Two long curved arrows map from the sheet on the left to the grid on the right: the upper arrow labeled Llm lands on z′; the lower arrow labeled SipIT lands on z.

Centered below: δ > 0 ⇒ ε > 0.
29 1
very linear time
Method
Mean Time (s)
Accuracy
HARDPROMPTS
6132.59土104.61
0.00
BRUTEFORCE (ours)
3889.61 ‡ 691.17
1.00
SIPIT (ours)
28.01 土35.87
1.00
8
29 likes 1 reposts

More like this

×