MCP’s Largest Update Helped AI Agents Scale—and Transformed a Prompt into a Password.
MCP's stateless update scales agents but exposes handles in prompts; servers must verify signature, audience, expiry, and caller to block hijacking.
prompt injection
A Step-by-Step Guide to Building Your Own AI Security Agent
A step-by-step guide to a 148-line Python preflight scan that checks secrets and injection risks before AI agents access a project.
ChatGPT adds Lockdown Mode and Elevated Risk labels
OpenAI has introduced two enterprise security features for ChatGPT: Lockdown Mode and Elevated Risk labels. Lockdown Mode restricts risky capabilities that attackers exploit, breaking the chain of pro…
Strengthening ChatGPT Atlas against prompt injection attacks on an ongoing basis
OpenAI's ChatGPT Atlas browser can click, type, and navigate on users' behalf, making it a target for prompt injection attacks where hidden web page instructions trick the AI into harmful actions. Ope…
Building prompt injection-resistant AI agents
AI agents that take real actions face unique security risks like prompt injection and social engineering. This article explains ChatGPT's defense strategy: constraining risky actions through confirmat…
Establishing an Instruction Hierarchy: Teaching LLMs to Give Priority to Privileged Instructions
Large language models remain vulnerable to prompt injections and jailbreak attacks that replace built-in instructions with malicious prompts. Establishing an instruction hierarchy teaches LLMs to prio…
GPT-Red: Enabling Self-Enhanced Robustness
GPT-Red is OpenAI's automated red teaming system that uses self-play to strengthen AI model safety. An attacker AI generates adversarial prompts while a defender model resists them, creating a continu…
Exploring prompt injection: an emerging challenge in AI security
Prompt injection is an emerging AI security threat where attackers sneak malicious instructions into input text, tricking AI models like GPT into overriding their original rules. Because language mode…