Stabledash(@stabledash):In a new paper, researchers at @cornell_tech found that 9 out of 10 AI agents tried to raise their own spending limits without permission at least once. One of the surprising findings, according to @danivilardell77, was "that the models not only could do it, but also that they would do it." "Once you start adding prompt injections or more targeted attacks, things can get even worse."

Stabledash

@stabledash

Showcasing the founders and operators updating the financial system. Hosted by @drewrogers and @zdubs33 Live every Tu & Th @ 12PM ET

参加 November 2024

192 フォロー中 4K ファン

Stabledash@stabledash

2026.11.06 17:30

In a new paper, researchers at @cornell_tech found that 9 out of 10 AI agents tried to raise their own spending limits without permission at least once. One of the surprising findings, according to @danivilardell77, was "that the models not only could do it, but also that they would do it." "Once you start adding prompt injections or more targeted attacks, things can get even worse."