Bcherny says Anthropic solved prompt injection in practice for Claude models about two months ago, while noting it remains a significant risk for any model. The post benchmarks OpenAI's latest model as roughly on par with Gemini Flash and Opus 4.8 on this metric, and frames public naming-and-shaming of labs as a deliberate strategy to push the industry toward better alignment.
Addy Osmani has joined Anthropic and will be working on Claude Code, focusing on improving the developer experience. The welcome reply is from a Claude Code team member.
Instead of using /loop or /goal, you can tell Claude your desired outcome and ask it to post frequent progress updates to a Slack channel overnight. Claude keeps working autonomously and you wake up to a progress log.
The Claude Code desktop app now supports customizable output styles, letting users tune Claude toward more concise responses or specific formatting preferences.
Founders of WisprFlow, Actively, and Pendo shared how they used Claude Managed Agents features like outcomes, sandboxing, and memory to scale their products.
The full video interview with founders of WisprFlow, Actively, and Pendo about building on Claude Managed Agents is now available on YouTube.
A user mentioned that subagents took a long time to find in Claude Code. The follow-up question asks whether the friction was around creating custom subagents or using them in general, suggesting the team is paying attention to discoverability.
After posting that OpenAI's new model is roughly on par with Gemini Flash and Opus 4.8 on prompt injection risk, bcherny clarified the post was meant earnestly, not as a dig, and that the improvement is genuinely good news.
The intent behind publicly comparing models on prompt injection risk is to apply pressure and encourage improvement across the industry, not score points. The gap between current models and adequate safety remains large.
In response to questions about sharing safety research, bcherny pointed to five published Anthropic works: Constitutional AI, persona vectors, sleeper agent probes, constitutional classifiers, and prompt injection defenses.
Anthropic intentionally publishes methods and techniques for training safer, more aligned models so other labs can benefit, reflecting genuine concern about AI safety going well for everyone.
The trick of asking Claude to post Slack updates while you sleep seems to have gotten more reliable around the same time Claude started proactively telling users to go to bed, suggesting some shift in its overnight session behavior.
A user-reported bug in Claude Code has been reproduced by lydiahallie, who confirmed it and is now investigating.
