​⚡ Who can build the best hacks and defenses for LLMs? ⚡🤖

​What if LLMs had hidden motives? Models deployed in the U.S. government and the world's most important companies might secretly be looking for opportunities to sneak backdoors into code, or otherwise pursue nefarious objectives. Could we catch these model hacks? This weekend, we're playing a capture the flag game with two teams:

• ​​​The red team: The red team will create realistic examples of models that hide secret objectives (e.g. the model secretly tries to share data with attackers)

• ​​​The blue team: The blue team will try to detect these objectives.

​We'll bring the compute and some examples to build on. You bring the research hustle.

​More details here: http://bit.ly/3JlJ3Q2

​Hosts

• ​Redwood Research (redwoodresearch.org)

• ​Constellation (constellation.org)

​​🗓️ Schedule (Subject to Change):

​​Sept 13th 2025

• ​​9:30 AM: 🚪 Doors Open

• ​​9:45 AM: 🎤 Opening Remarks

• ​​11:00 AM: 💻 Start Coding!

• ​​12:00 AM: Lunch Provided

• ​​6:00 PM: Dinner Provided

​​Sept 14th 2025

• ​​9:30 AM: 🚪 Doors Open

• ​​12:00 PM: Lunch Provided

• ​​2:00 PM: 📤 Project Submission Deadline

• ​​5:30 PM: 👩‍💻 Awards ceremony

​​➡️ Next steps ➡️

1. ​​Join the discord: https://discord.gg/7qEDgCWAwm​👀 #teamsearch is where you can look for teams

2. ​​Register on Devpost: https://redwood-af.devpost.com/

​​Notion: http://bit.ly/3JlJ3Q2