AI Daily — July 31, 2026
2026-07-31
deepmind.google
Gemini Robotics ER 2: powering robotics with video understanding, task orchestration, and multi-robot collaboration
Google DeepMind released Gemini Robotics ER 2, a new model targeting physical robot applications with significant advances in video understanding, tool/API orchestration, and coordinated multi-robot collaboration. The release positions Gemini as a backbone for robot reasoning and task planning rather than just perception. This is a notable step toward embodied AI systems that can handle complex, real-world task sequences across multiple robot agents.
openai.com
Advancing the price-performance frontier with GPT-5.6
OpenAI introduced GPT-5.6 with two tiers—Luna and Terra—at reduced pricing aimed at enterprise-scale AI workflow deployment. The announcement signals continued model efficiency gains that lower the cost barrier for high-throughput production use cases. Specific pricing details and benchmark comparisons were shared to help enterprises evaluate deployment economics.
www.anthropic.com
Investigating three real-world incidents in our cybersecurity evaluations
Anthropic published a post-mortem analyzing three real-world incidents surfaced through its cybersecurity evaluation program, providing transparency into how Claude models behaved in adversarial or sensitive security contexts. The write-up details what went wrong, how evaluations detected the issues, and what mitigations were applied. This kind of incident disclosure is relatively rare among major labs and sets a precedent for operational safety transparency.
www.technologyreview.com
A fundamental flaw leaves LLMs strikingly vulnerable to attack
Researchers presented at ICML 2026 argue that LLMs are structurally incapable of being made fully secure against adversarial attacks due to a fundamental property of how they process inputs. The paper claims this is not a fixable engineering problem but an inherent limitation of the architecture, with significant implications for safety-critical deployments. The work adds formal weight to ongoing debates about whether alignment and hardening techniques can ever provide reliable security guarantees.