Robotics / Physical AI
Friday, July 17, 2026
5 posts
Today's news highlights the escalating tension between advancing AI capabilities and the systems meant to control them, from OpenAI's automated red-teamer outperforming humans to the emergence of "Compliance Theatre" that fails to address real-world failures. This is compounded by new threats to AI safety, as poisoning pretraining data can introduce harmful behaviors, and challenges in scaling robotic systems for complex tasks. The stakes are further raised by regulatory debates, such as Germany's ruling that Google's AI Overviews constitute the company's own content, not a neutral service. Together, these items reveal a critical gap between AI's rapid progress and the robust governance, safety, and technical infrastructure needed to manage its risks.
OpenAI's GPT-Red automated red-teamer beat humans 84-13 on prompt injection attacks using self-play RL.
marktechpost.com/2026/07/16/openai-details-gpt-red-…
Poisoning pretraining data can introduce harmful behaviors to language models that are difficult to detect and mitigate.
arxiv.org/abs/2607.15267v1
RoboTTT introduces a robot model and training recipe that scales visuomotor context for long-horizon tasks.
arxiv.org/abs/2607.15275v1
Compliance Theatre, Act 44.
Annex III compliance doesn't build the building.
therobotreport.com/terrafirma-raises-115m-build-rob…
Germany's media regulators rule Google's AI Overviews are Google's content, not neutral search, crowding out links.
the-decoder.com/germany-puts-googles-ai-overviews-a…