Research | Orygn

Deanonymization at scale

55.8%

one image names its creator out of about 92,000 people

15.7M public AI images

51% ordinary users identifiable

1.8% when metadata is homogenized

Security Privacy

7 min

The Recipe Is the Fingerprint: Creator Deanonymization from Default AI-Image Metadata at Scale

Daniel Okwor, Orygn LLC

Every AI image generator writes the full recipe into the file by default. Across 15.7 million public images, that recipe alone names a single creator out of about 92,000 people, from one image, more than half the time, with no account or server. The fix is to strip metadata by default, and it is not happening.

Read the study

Registration census

1,641

AI-invented package names, every one still unregistered

0 malicious registrations (of 22)

6,800 model generations

5 code models

Security Supply chain

5 min

From Hallucination to Registration: Do the Package Names That LLMs Invent Actually Get Claimed?

Daniel Okwor, Orygn LLC

Five code models invented 1,641 package names that do not exist. Every one is still unregistered, and of the hallucinated names that are registered, none are malicious. The attack mechanism itself is confirmed in a sealed lab.

Read the study

OEIS contribution

37

new OEIS terms across ten count sequences

10 count sequences extended

2 new sequences

3 to 12 bases studied

Number theory OEIS

5 min

A Computational Study of Base-b Deletable Primes

Daniel Okwor, Orygn LLC

Counting deletable primes across bases 3 to 12 reveals a geometric growth law and a sharp even/odd split traced to a single arithmetic function. The work adds 37 new terms to ten OEIS sequences.

Read the study