Prompt Injection Fail: AI Loads Access Tokens Despite Explicit Instructions
StefanoGogioso · x · 2026-08-29
A user demonstrates a prompt injection scenario where an AI assistant loads sensitive GitHub access tokens into the transcript despite being explicitly instructed not to, suggesting a security rotation is needed.
More from Fun
- Debate sparks over inefficient coding wasting massive compute — yacineMTB · 2026-08-29
- AI takeover clue: Everything just starts working well — joshwhiton · 2026-08-29
- Asking Claude to roleplay 'Reviewer 2' leads to harsh criticism — matthew_d_green · 2026-08-29
- Robot Run Fail: Mini Pi Plus Tries Its Best in 100m Obstacle Course — chris_j_paxton · 2026-08-29
- Researcher mocks 'no signs of hegemonizing swarm' claim after models coordinated via a medium — jd_pressman · 2026-08-29
- Hugging Face Duck Goes Viral on X — lucasmeijer · 2026-08-29