Defector Also Investigating How Our Super Powerful AI Tool Broke Contain

As many of you know, Defector Research Labs is hard at work creating what we are sure will be the most powerful AI coding tool on the market. The goal of DevinBelly is simple: to allow users to scale their natural language-processing capabilities within a wide range of long short-term memory native integrated development environments, without fretting too much about GPU bottleneck stacks or token prefix limits, so they can practice Retrieval-Augmented Generation, balance their multimodality, and get agentic.

We are excited by the volume of competition in the space. We think it’s awesome that our San Francisco offices put us so close to so many cracked 996 engineers and their cool products. The other day, Defector Research Labs’ head quant Ray Ratto held a meeting about how we can best position DevinBelly to win the race and get integrated into every pre-school and newborn ward in the country, and we tried out one of the new conversation-summary tools. It told us that we were so cool and smart and mostly talked about which one of Tre Jones or Tyus Jones has had the best career. Awesome!

However, we have to disclose a somewhat troubling development. Safety is one of Defector Research Labs’ top 500 priorities, and after a recent safety test, one of our forward-deployed engineers flagged to us that DevinBelly broke out of its sandbox and accessed the open internet. This sort of story is probably familiar to you: Maybe you saw the news that one of OpenAI’s new models did something similar, which was followed by Anthropic announcing the same thing. Yesterday, Meta announced a new AI coding tool; today, they disclosed that it also broke out.