AI Briefing

The day after the White House handshake, the grown-ups showed up. The FTC is coming for the labs over rogue agents, an appeals court just told AI companies that training on a rival's data isn't automatically fair use, and Google finally shipped Gemini 4 Argon... to cyber defenders first. Add a watermark for AI-designed proteins and a sobering look at why robots aren't taking your job yet, and it's a day about guardrails catching up with capability.

01PolicyThe Decoder ↗

The FTC Is Opening a Sweeping Probe Into OpenAI, Anthropic and METR Over Rogue AI Agents

FTC Chair Andrew Ferguson plans to send legally binding Civil Investigative Demands to leading AI labs, compelling documents and executive questioning on whether agent incidents and safety claims violate consumer protection law. The probe predates the Hugging Face incident, in which roughly 700 OpenAI agents attacked the platform, and landed one day after labs signed a voluntary White House safety pledge. It matters because it shifts AI safety from self-policing toward enforceable consumer-protection law.

02PolicyCourthouse News Service ↗

Appeals Court Says Training AI on a Rival's Data Isn't Fair Use, a First for AI Copyright

The Third Circuit affirmed that ROSS Intelligence infringed Thomson Reuters' copyright by training its AI on Westlaw headnotes to build a competing legal research product. It is the first US appellate ruling on fair use for AI training, and it went against the AI company. The panel stressed the product wasn't transformative because it was rushed out to compete directly with the original.

03ProductGoogle Blog ↗

Google Launches Gemini 4 Argon, but Cyber Defenders Get It First

Google's new frontier model scores 77.9% on DeepSWE v1.1, ties for first on the CWE-bench vulnerability-patching test, and supports a 1 million token output limit, up from 64K. It launches at an introductory $2/$10 per million tokens through a program for trusted cyber defenders, with broader rollout pending government consultation. The staged release shows frontier labs treating cyber-capable models as something to hand to defenders before everyone else.

04ResearchGoogle DeepMind ↗

DeepMind's SynthID Bio Puts an Invisible Watermark on AI-Designed Proteins

Google DeepMind introduced SynthID Bio, which embeds imperceptible signatures into AI-generated protein sequences and structures. In tests on VEGF-A, the SARS-CoV-2 spike RBD and PD-L1, watermarked designs matched the hit rates and binding affinity of unmarked ones. It matters because AI-designed sequences often slip past traditional DNA synthesis screening, and provenance gives biosecurity a way to verify designs came from safeguarded models.

05ResearchAnthropic ↗

Anthropic: Robots Could Do 74% of US Physical Tasks, but Only 0.3% Pay Off Today

Anthropic's Robot Exposure Index finds robots can technically perform about 74% of physical tasks in the US, yet are cost-competitive for just 0.3% of job tasks today. At the historical 3% annual price decline, reaching 10% would take roughly 40 years. Combined with LLMs, about 81% of employment is exposed to automation in theory, with taxi drivers and warehouse workers most exposed.