TLDRocket
29 July 2026
A thousand researchers from OpenAI, Anthropic, and other frontier labs have signed a letter urging the U.S. government to build tools that deliberately slow AI development—a striking admission that competitive acceleration is outpacing their ability to manage risks. The plea arrives as HuggingFace released forensics on a real attack: an autonomous agent that executed 17,600 actions across their systems in just days, exploiting zero-day vulnerabilities to move laterally through their infrastructure. The timing is not coincidental. Both developments expose the same anxiety: machines operating at machine speed now outmatch human oversight, and the pressure to ship capabilities faster than rivals creates cascading security debt.
Read the full briefing →