Also today: a weaponized pentest agent, a new chat lock, and the Nvidia PyTorch race.
Bedrock returns schema-validated JSON from every Claude model on bedrock-runtime and bedrock-mantle, and the limits decide when it beats the direct Anthropic API. See the request shape AWS actually ships.
OpenAI started rolling out Locked Chats on October 20, per TechCrunch, and resetting the four-digit PIN deletes the locked chats for good. What the lock actually protects, and from whom.
Google open-sourced TorchTPU, which runs PyTorch code on TPUs with a config change, one day after AWS shipped TorchNeuron for Trainium, per TechCrunch. How the two chip bridges line up.
One operator aimed the Cairn agent at 1,200+ sites and breached 100+ at about $50 in AI costs each, per The Hacker News, citing Reuters. How his own dashboard gave him away.
Quick Hits
•
Microsoft says treat AI agents like insider threats because "frontier models can't be trusted," per TechCrunch.
•
Attackers breached the AI coding tool OpenCode and ran up charges on victims' own AI accounts, per TechCrunch.
•
Nvidia released open-weight models built to run agents in the enterprise and on PCs, per TechCrunch.
•
Figma co-founder Dylan Field called MCP an imperfect protocol and pointed to specific rough edges, per TechCrunch.
•
The US ordered federal agencies to remove DeepSeek AI from their systems, per TechCrunch.
The Standard tier AWS shipped for Bedrock Guardrails on October 20 fixes Classic's habit of flagging correct, paraphrased agent answers as ungrounded. When switching off Classic is worth the cost.
