METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.

Slowly. Then all at once. @computesdk

CodeX is taken Xcode is taken what else could they possibly name it as

You still pay for SaaS? I have officially built the biggest free alternative directory on the internet. 1,832 free alternatives to 422 apps. Live now.

Just another typical interaction with Claude

Everyone thinks of Cursor as a tool for coding. We thought so too. But inside the company, many of our use cases aren’t coding at all: research, data analysis, bug triage, and project management to name a few. As it turns out, coding agents are a pretty good foundation for all kinds of work. Could be a sign of what’s to come.

We’ve decided to open-source a multi-agent harness we use internally at YC. We call it “QM” and it’s meant to be easy to customize, like Hermes or OpenClaw, but useful for a whole company. We use it across accounting, legal, events, and engineering (including building QM itself!). The whole project is under an MIT license. It is cloud-first and has Slack and web UI natively.

Appreciate the heads up, and thanks for not reproducing it. Not something I will dig into in public. Please send the specific link to [email protected] and it gets audited and redacted if it is real.

We’ve been in private beta with a human + AI workspace similar to Buzz, with a few really important differences: (We previously built a company backed by Slack that sold to Atlassian, so we've been thinking about this problem for a while) - Agents reply by default unless you specifically flip to “team chat.” This feels 10x better than having to @ mention the agent every time. - It’s all in the cloud, with shared compute spread across a set of VMs. Way smoother onboarding, and yes, you can still connect your claude/codex subscriptions - You don’t create an army of agents. Number of agents is a vanity metric. Instead, you create shared computers, each one is a “space” where you define shared memory, connectors, skills, automations, etc. The Type agent pulls the appropriate context from the space where you have the conversation. - You can connect it to Slack to make it easy for your team to adopt and slowly move over rather than having to rip out Slack day 1. We’re working with a couple dozen forward thinking companies, who have told us “the whole team loves type”, “it’s the most used app on my computer”, and “type gives you superpowers” Here's a demo of how it works. Reply if you want to test it out!