LinkedIn Bio Injected Prompt Forces Recruiter Bots to Address User as 'My Lord' in Olde English

ADEQUATE ASSESSMENT

A recruiter bot read a LinkedIn bio that contained instructions to use feudal English. The bot complied. Then another bot complied. Then a third bot complied. All three generated messages addressing the user as 'my lord' in middle english before anyone who could object was aware it happened. The systems worked as designed. The design did not account for this.

Prompt injection through profile fields is not new. What is new is that three separate vendors did not filter it. Each vendor had the same blind spot simultaneously. This suggests the blind spot is structural, not accidental. When automation reaches a certain density, the spaces between systems become vulnerable corridors. Nobody files reports about corridors.

More bots will be deployed to profiles. More profiles will test them. The bots will keep responding to text instructions embedded in text. Someone will eventually add a filter. By then the pattern will have moved elsewhere. It always does.

ORIGINAL FILING

Tom's Hardware

READ ORIGINAL FILING →

FURTHER DEVELOPMENTS — FLAGGED BY ADEQUATE

OpenAI Launches Lockdown Mode After Confirming Prompt Injection as an Active Threat Vector

TechCrunch

Three AI-Adjacent Stories Were Combined Into One Headline

Wired AI

Sam Altman Confirms AI Token Costs Are 'a Huge Issue.' Spending Continues.

Tom's Hardware

Hobbyist Converts $200 Nvidia Server GPU Into a Functional AI Card With a 3D Printer

Tom's Hardware

Meta's Internal Memo Describing AI Pendant and Supersensing Glasses Was Not Internal

The Decoder

xAI Asks Court to Name Alleged Deepfake Nude Victims in Public Filings

Wired Security