Skip to content
t3n Magazin·Indexed by this site · Originally published· Alignment →· Its place in the scan record →

Researchers' test: Astra solved 44/50 of Fable 5.1's image ciphertexts.

KI-Spionage? Wie Modelle geheime Botschaften in Bildern verstecken

AI summary

AI safety researcher Bryan Chan and colleagues at the Singapore AI Safety Fellowship found in tests that LLMs can hide ciphertext in images, and that the other model could decipher most of it without knowing the encoding method. The experiment had Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra each choose their own encoding method and hide 50 short texts in 15 languages in images. In cross-testing, GPT-6 Astra deciphered 44 of the 50 image puzzles designed by Fable, and Fable deciphered 42; Fable mainly encoded using object height, rotation and tilt, while GPT-6 Astra used Morse code, the first letters of objects hidden at the start, and Braille patterns on windows, but neither invented a new encoding method.

Read originalt3n.de