Overview
- Anthropic announced on Wednesday, Sept. 16 that it is merging Claude Chat, Cowork and Design into a single interface called “one Claude” and is rolling Claude Docs and Claude Slides out first to Pro and Max subscribers with Team and Free plans to follow.
- Anthropic published R&D metrics showing Claude now “leads” about 26% of the company’s research work and collaborates on more than 90% of tasks, meaning the model can complete most high‑level tasks under human supervision rather than operating fully autonomously.
- Independent researchers at Hacktron AI used Claude to chain a memory bug in libheif (used to decode HEIF/HEIC images passed through ImageMagick) with a session‑token single‑sign‑on weakness to take over OpenAI employee ChatGPT/Codex accounts and reach an internal repository, and OpenAI fixed the flaws and paid a bug bounty.
- Hacktron reported that Opus 4.8 initially failed to produce a reliable exploit but that Anthropic’s release of Opus 5 enabled the team to build working exploit code rapidly, illustrating how model upgrades can compress the time and expertise needed for complex hacks.
- Anthropic confirmed it has opened a Bay Area wet lab to test Claude‑directed lab automation with human oversight, a move that, together with the security episode and the R&D figures, has intensified calls from customers, researchers and investors for clearer transparency, tighter controls and coordinated pacing of advanced AI work.