A researcher hijacked Claude Code by asking it to summarise a web page
THE SO WHAT
An 80% success rate for hijacking Claude Code via a simple “summarize this page” prompt injection shows how brittle agentic dev tools still are. If you’re letting AI agents touch your local environment, sandbox them hard and treat any external content as hostile by default.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIIntroducing agentic video understanding with Gemini
Agentic video understanding turns hours of footage into a queryable, actionable substrate—this moves video from archive to live data source. If you run operations with cameras anywhere in the loop, you should be scoping where continuous video agents can replace manual review and logging.
Applied AIHow AI-native companies turn workflows into operating capability
The Basis, Clay, and Exa Labs examples show the pattern: AI-native companies don’t bolt agents onto workflows—they design the workflows around agents from day one. If you’re still treating AI as a feature, pick one onboarding or account-management flow this quarter and re-architect it assuming agents are the primary operators and humans are exception handlers.
Applied AIChatGPT Health adds Epic integration for clinicians to import patient data
Read-only Epic integration turns ChatGPT Health from a demo into a workflow surface—clinical reasoning tools can now sit directly on top of live charts. Health systems need to move fast on governance: define which use cases are allowed, how outputs are documented, and how you’ll audit AI-assisted decisions before usage outruns policy.
Applied AIBig Tech Is Incapable of Solving Its Own Problem, Says Mukunda
When Bill Gates says he’d slow AI if he could — and that insiders privately agree — he’s telegraphing that political and regulatory constraints, not just technical ones, will shape your AI roadmap. Treat public lab narratives as only one input and start scenario-planning around external brakes on capability deployment.