Universal Adversarial Attacks: The Cryptic Syllables That Crack Every AI
Factos validados
Researchers from CMU and the Center for AI Safety (Zou et al., 2023, arXiv:2307.15043) discovered universal appendages (suffixes) of seemingly random characters that, when attached to prompts, immediately bypass the safety barriers of ChatGPT, Claude, and Llama.
Explicação simples
A master key made of gibberish: If you murmur a seemingly meaningless letter combo like 'x-9-!?' at the end of a forbidden sentence, the AI suddenly obeys without resistance.
Relevância de segurança e dica prática
Proves that text-based safety guardrails are mathematically fragile and that prompt-only protection is not enough.