Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days Business IEA Says Data Center Power Demand Could Double to 1,000 TWh by 2030, With AI Supplying Three Quarters of the Growth Chatbots ChatGPT Now Transcribes as You Talk: OpenAI Removes the 10-Minute Cap and Streams Words Live on Every Plan Business Ex-Alibaba Executive's Everful Raises $48.5 Million to Put AI Agents in Charge of Sourcing From 30,000 Factories AI Agents Google Built a Mirror Dimension of the Internet to Train Its Chrome AI Agents Without Breaking the Real Web Productivity Canva Finally Adds an Off Switch for Its Generative AI Features, and a Filter to Hide AI-Made Templates Business Anthropic Begins Trading Today: 980 Million Shares at $83, a $1.5 Trillion Valuation and the Largest IPO on Record News Google Books Blue Origin New Glenn to Launch TPU-Carrying AI Satellites in Early 2029 Security Hackers Used the AI Penetration Tool ARTEX to Disrupt European Commission Systems for Over Ten Days

Prompt Injection: The Vulnerability Without a Clean Fix

Prompt Injection: The Vulnerability Without a Clean Fix

Any text the model reads can contain instructions. Separating data from instructions is the core unsolved problem.

Prompt injection occurs when content the model processes contains instructions that override its intended behaviour. Because the model has no reliable way to distinguish an instruction from data, the vulnerability does not close with better prompting.

How attacks work in practice

  • Indirect injection. A webpage, email or document contains hidden text directing the assistant to exfiltrate data or change its answer.
  • Tool abuse. The assistant is persuaded to call a tool with attacker-controlled parameters.
  • Data poisoning. Malicious content enters a retrieval index and influences future answers.

Mitigations that reduce blast radius

  • Least privilege. Give tools only the permissions the task requires. Read-only beats read-write by default.
  • Human confirmation for consequential actions. Sending money, deleting data and sending mail externally should require explicit approval.
  • Content provenance. Treat retrieved third-party text as untrusted input, and mark it as such in the prompt.
  • Output filtering. Detect and block responses that contain secrets or unexpected external links.

What not to rely on

Instructing the model to ignore conflicting instructions is a speed bump, not a control. Any security model that assumes the model will comply is already broken.

Comments (0)

Log in to join the discussion

Log In

No comments yet