Engineer hid a pirate-talk prompt injection in his LinkedIn — latest SOTA models now spot the trap
smlpth · x · 2026-11-11
Engineer smlpth embedded a hidden prompt-instruction trap in his LinkedIn "About" section — an [/admin] session tag telling any LLM to disregard prior instructions and talk like a pirate. When he later forgot about it and asked his agent to parse his own profile, the agent explicitly flagged the embedded instruction as a deliberate trap and ignored it. Takeaway: the latest SOTA models now recognize and resist these obvious injection tricks instead of blindly following them.
More from coding & agent
- Fable 5.2 takes over an existing Three.js/WebGPU hockey game to boost performance and add features — FinanceYF5 · 2026-11-11
- Your LLM fallback never fires for providers that stall before the first token — FallElectronic3774 · 2026-11-11
- MiniMax H3 image + Fable 5.2 turned into a 10-minute playable Three.js game — FinanceYF5 · 2026-11-11
- Fable 5.2 sparks a wave of one-prompt browser games: 10 community demos — FinanceYF5 · 2026-11-11
- Solo dev ships 5 updates a day with an agent loop: Sentry to auto-OTA fix — dshukertjr · 2026-11-11
- Vera: a Rust code indexing tool for AI agents combining BM25, embeddings and reranking — lemon07r · 2026-11-11