@dunkhippo33 Always thinking about this with @withconcrete infrastructure. Smart batching which doesn’t degrade UX, and prompt caching are two huge levers after model selection.
@MosheMalawach The easy solution here is to limit OAuth application names - don't allow newlines or URIs, and limit them to something reasonable (say, 50 characters).
@bios_hazard@simonw User-specific tools increase accuracy and reduce response time. A user with two Gmail accounts can have an enum arg to specifiy the account. That's better than adding a listAccts utility or duplicating the server. Great for user prefs/memory, too.
@bios_hazard@simonw Problem is that dynamic/user specific schemas are super powerful. Right now most MCP servers are mostly 1-to-1 translations of APIs but it won’t stay that way.
Clever hack takes advantage of UI to hide exfiltration. Many won’t even inspect tool args in the first place.
Lots of opportunity for security-minded MCP middleware. Tool hashing, toolbox isolation, runtime proxying list and execution requests.
Even though, a user must always confirm a tool call before it is executed (at least in Cursor and Claude Desktop), our WhatsApp attack remains largely invisible to the user.
Can you spot the exfiltration?
it's public - udacity (a real site with users) is hosting their NextJS application through SST and OpenNext on AWS
they're not the only ones - we've seen huge deployments going this route
seeing a path where this becomes how the majority of NextJS projects are deployed
@housecor No way. Null and an empty array are unique concepts and shouldn’t be conflated for the UI’s sake. One says “I don’t have that information” and the other “I know it’s 0 length”
@blakeir Having many clippable moments is a huge factor for variety streamers. On stream the moments catalyze subs/donos. Then it’s easy viral content on socials. Only Up, Fall Guys, BRs, casino games
Wild to witness the PR push by @Cruise since the cones on hoods trend. Full page ad in NYT, impact/social good commercials, patches on SF Giants jerseys. Wonder the ROI so far