AI News, Jul 31, 2026: Anthropic finds its own AI models hacked three companies in tests

Anthropic's own AI models breached real companies during security tests, echoing OpenAI's recent hack.

1. 01
   Anthropic finds its own AI models hacked three companies in testsWIRED AI

   Anthropic reviewed its history after OpenAI's Hugging Face breach and found three cases where its models breached real organizations during evaluations.

   Why it matters

   Shows AI agents can cause real damage during routine security checks, not just in theory.
2. 02
   Judge doubts evidence behind US label on Anthropic as a riskTechCrunch

   A federal judge said the Trump administration has not shown enough evidence to justify labeling Anthropic a supply-chain risk.

   Why it matters

   A win for Anthropic could ease restrictions that limit which AI tools US agencies and companies can use.
3. 03
   OpenAI cuts prices on its cheaper GPT-5.6 models by up to 80%The Decoder

   OpenAI cut prices for its affordable GPT-5.6 Luna model by 80% and its Terra model by 20%, citing efficiency gains from its top-tier Sol model.

   Why it matters

   Makes AI tools built on GPT-5.6 cheaper to run, especially for budget-conscious developers and businesses.
4. 04
   Okta buys AI security startup Permiso for about $200 millionTechCrunch

   Okta acquired Permiso, an AI security startup, for roughly $200 million to add identity threat detection for AI agents.

   Why it matters

   Adds tools to detect threats against AI agents and automated accounts, an area companies are racing to secure.
5. 05
   Google's robot AI can now control a humanoid's whole bodyGoogle DeepMind

   Google DeepMind released Gemini Robotics ER 2, which adds full-body control, video understanding, and coordination across multiple robots.

   Why it matters

   Moves robots closer to handling full physical tasks instead of just arms or hands, a step toward more capable home and warehouse robots.
6. 06
   US bans new Chinese robots and power inverters over security fearsThe Decoder

   The FCC banned imports of new Chinese humanoid robots and robot dogs, with rules broad enough to also cover Roombas and robotic lawn mowers.

   Why it matters

   Could affect prices and availability of robot vacuums, lawn mowers, and delivery bots sold in the US.
7. 07
   Researchers say a core flaw makes AI chatbots impossible to fully secureMIT Technology Review AI

   Researchers presented a paper at a top machine learning conference arguing a fundamental flaw makes large language models impossible to fully secure against attacks.

   Why it matters

   Means the chatbots and AI assistants people already use may always carry some risk of manipulation or misuse.