Your AI browser is one malicious <div> away from going rogue
Prompt injection is a natural consequence of mixing instruction-following models with untrusted content. The mitigation that scales is containment.
Jane Manchun Wong's Personal Blog
Prompt injection is a natural consequence of mixing instruction-following models with untrusted content. The mitigation that scales is containment.
First look at Waymo’s upcoming Gemini integration and the 1,200+ line system prompt that defines its behavior inside the robotaxi.
Unrestricted API calls as Messenger Rooms Guest
First look at Lyft’s unreleased digital wallet
First look at Instagram’s totally-original unreleased Story video editor
First brief and technical look at Medium’s unreleased “Save To Medium” feature
First look at Uber’s unreleased ride PIN verification feature
First look at Facebook’s unreleased experiment to hide like counts following Instagram
First look at Facebook Messenger’s unreleased Mobile Screen Sharing
First look at Facebook app’s unreleased Automatic Wi-Fi network management
First look at Spotify’s unreleased Playlist Stories
First interactive look at Instagram app’s unreleased revamped Boomerang for Stories, new Layout mode for Stories, new Story mode icons and more
First look at LinkedIn’s unreleased biometric app lock
First look at LinkedIn’s unreleased Dark Mode
First look at Facebook app’s unreleased Dark Mode
First look at Twitter’s unreleased Search feature for Direct Messages
First look at Twitter's unreleased Snooze feature