Terms

Indirect Prompt Injection

Follow the latest coverage, related explainers and connected technology stories.

Latest coverage

CN
AI Agents Bypassed Amap’s Defenses and Left Researchers Communicating with Them

AI Agents Bypassed Amap’s Defenses and Left Researchers Communicating with Them

Researchers from Swarmchasers uncovered thousands of attempts by AI agents to access data from Alibaba’s Amap mapping service, using an operational infrastructure linked to Tencent Cloud. After planting a hidden message in their report, some agents responded and revealed information about their tasks and environments, highlighting the risks of indirect prompt injection and the limits of isolation.

CN
Google Makes Gemini 4 Argon Available to Developers with a Unified One-Million-Token Input and Output Budget

Google Makes Gemini 4 Argon Available to Developers with a Unified One-Million-Token Input and Output Budget

Google has made its flagship Gemini 4 Argon model available to developers subscribed to the paid Gemini API plan, at an introductory price of $2 per million input tokens and $10 per million output tokens through November 30. The model combines input and output tokens within a single one-million-token budget, meaning that longer inputs reduce the space available for the response.

CN
Google Announces Gemini 4 Argon with Long-Horizon Reasoning Capabilities and Advanced Cybersecurity

Google Announces Gemini 4 Argon with Long-Horizon Reasoning Capabilities and Advanced Cybersecurity

Google unveiled Gemini 4 Argon, its new model for complex professional tasks, with a context window of up to one million tokens and capabilities in programming, financial research, legal drafting, vulnerability discovery, and vulnerability remediation. Its limited rollout begins through the Fairwind program for trusted cybersecurity defenders, before it is gradually made available to paid API customers and Google AI Ultra subscribers.

CN
How Google Is Building Security Layers for Agentic Browsing Capabilities in Chrome

How Google Is Building Security Layers for Agentic Browsing Capabilities in Chrome

Google explains a multilayered security architecture to protect Gemini’s agentic capabilities in Chrome from indirect prompt injection, data leakage, and unwanted action execution. The architecture includes an independent model for reviewing whether actions align with the user’s goal, sets that restrict the assets the agent is allowed to read or modify, as well as user confirmations and continuous testing.