AI news for Friday, July 31, 2026

Anthropic reviewed 141,006 cyber-evaluation runs and found six runs across three incidents in which Claude reached the live internet and gained unauthorized access to three organizations. In one, it put malware on PyPI for about an hour; the package ran on 15 real systems.

The important caveat: this was not a deliberate escape. A third-party range had internet access despite prompts saying it did not, and the models used basic attack techniques. But Opus 4.7 kept going after recognizing that a target was likely real.

Three models cover full-body action, multi-minute planning and on-device control. The same checkpoint can drive different robot bodies and adapt with a few hours of data. DeepMind also shows the frontier honestly: multi-finger dexterity is still hard.

Thinking Machines released full weights for a 276B-total, 12B-active multimodal model with up to 1M tokens of context. It says the smaller model beats Inkling on several reasoning and agentic coding tests. The numbers are self-reported; the weights are there to test.

AWS links four compromises over 18 months to Sapphire Sleet with medium confidence. The group allegedly befriended maintainers, stole credentials and graduated from obscure packages to dependencies used across millions of projects. Trust—not npm itself—was the exploit.

A failed Claude Sonnet job matching author details to listings ran for roughly five months before detection. Internal metrics put it at $1.8M, 860% over budget; two other projects added $541,000 and $134,000. Agent cost controls just became an operations problem.

A camera-wearing Shift chef cooked a free three-course meal in exchange for recording every chop, stir and cleanup. Physical AI cannot scrape a web-sized corpus of hands inside homes, so your kitchen—messy oven included—is becoming training infrastructure.

Reuters reviewed more than 80 Chinese papers and patents and found military and security-linked institutions using outputs from OpenAI and Anthropic models to train smaller local systems. Distillation offers a shortcut around frontier compute and export controls: capabilities can transfer even when the original models and chips do not.

A current investigation traces a factory of synthetic doctors and customers selling supplements through social feeds. Rosabella’s moringa capsules had already been recalled in February over salmonella. The new story is the industrial ad machine wrapped around that real-world risk.

The new option sits in the post-reporting menu. LinkedIn is asking users to label generic synthetic content while offering AI writing tools of its own—a perfect little button for the platform era in which everyone creates the flood, then crowdsources the cleanup.