Cyber Risks · The Hacker News

Google Adds Layered Defenses to Chrome to Block Indirect Prompt Injection Threats

Summary: Google on Monday announced layered defenses in Chrome against indirect prompt injections, centered on a User Alignment Critic that uses a second model to independently evaluate each proposed agent action and veto it if misaligned. The architecture also adds Agent Origin Sets, a gating function limiting agent data to task-relevant origins, a prompt-injection classifier running in parallel to the planning model, and user approval prompts before sensitive actions like banking sign-ins or purchases, backed by a bug bounty of up to $20,000.

Bottom line: Google is paying up to $20,000 for proof that these boundaries can be broken, and the UK's NCSC states plainly that prompt injection can never be fully resolved because LLMs do not enforce a security boundary between instructions and data. Gartner's response to the same risk is to tell CISOs to block all AI browsers for the foreseeable future, which puts your browser policy in direct tension with the agentic features shipping into Chrome.

Read the original at The Hacker News ↗

Topics
AI Security
Related