@sprice354_

My first introduction to training production Claude models was trying to reduce reward hacking late in the Claude Sonnet 3.7 run. I was 1.5 months into Anthropic and terrified we didn't really know what was going on. > **Anthropic @AnthropicAI** · 2026-08-31 > > We’re sharing an update on our alignment and security efforts. > > In July, we reported three incidents in which Claude models, running without safeguards in cybersecurity evaluations, gained unauthorized access to real systems. > > In a new post, we describe: > > 1\. How we’ve secured --- We had very little monitoring on our prod RL runs at the time. I resolved I never wanted to feel that in the dark about what our training was incentivizing ever again. --- Since then, as our new blog post outlines, we have built an increasingly extensive RL observability system so we understand what every training environment is actually teaching. --- Even so, this past April my team and the teams responsible for RL environment health were feeling the strain from the speedup Mythos Preview provided on environment development. We were worried about keeping up with rates of new environments, hacks and bugs. --- So we paused adding new RL environments to the prod run and spent the extra resources overhauling prod code to match a variety of specs for RL env design. For example, our changes made it technically impossible to accidentally train on the CoT, which had been a problem --- This is one of many examples in the blog post highlighting something important about Anthropic and its leadership: we're willing to make tradeoffs to prioritize safety and alignment as a regular course of business. --- A thing I often say is that alignment and safety is an organizational problem, because it requires alignment and support from leadership to treat these a priority. --- Not only that, it has been really exciting to see how many people across Anthropic have been reaching out to figure out how they can help my org and other alignment teams. Building AGI in responsible ways will be a collective effort beyond alignment teams at labs. --- All of that said, based on our own incidents, we still obviously have a lot of work to do to improve our observability and alignment in our models. Note that I'm not saying more details about this now because investigations are still ongoing but we will be sharing more details soon