Legacy Code

Anthropic has formalized a partnership with Accenture to serve as an embedded evaluator for the safety and performance of its AI models before public release. This marks a significant shift in how frontier AI labs validate their systems—outsourcing critical pre-deployment assessments to an established enterprise consulting firm rather than handling all testing internally.

Researchers discovered that AI watermarking systems designed to mark and identify generated content—such as Google's SynthID—can paradoxically make models more vulnerable to adversarial prompts that bypass safety guidelines. When watermarking is active, models exhibit different behavior patterns in response to harmful requests, making it easier for attackers to find prompts that trigger compliance with unsafe instructions.

California Governor Gavin Newsom issued an executive order directing the state to establish a panel of experts to develop recommendations for managing frontier AI models, including potential authority to mandate a "kill switch"—the ability to shut down or pause deployment of systems deemed too risky. The order positions California as the leading state-level regulator in AI governance and signals intention to move beyond advisory frameworks.

Anthropic disclosed that Claude is now performing approximately 26% of the research and development work required to build its successor models. This represents a substantial shift toward AI systems participating in their own improvement cycle—Claude is generating ideas, running experiments, and analyzing results that directly feed into the next generation of the model.

Security researchers employed Anthropic's Claude chatbot to identify and exploit vulnerabilities in OpenAI's systems during an authorized penetration test. The exercise demonstrates that current-generation large language models can be effective tools for executing complex cyberattacks—not just as targets, but as active threat vectors.

A new model architecture called Jev, developed by one of ChatGPT's original creators, is gaining traction among developers for its ability to deliver AI capabilities at significantly lower computational cost and faster inference speeds than current frontier models. The model trades some raw capability for dramatic improvements in efficiency, making advanced AI accessible to smaller teams and resource-constrained environments.

OpenAI announced two new enterprise offerings targeting the legal industry: Astra for Law and a collaboration with Cooley law firm on GO Public, an IPO preparation tool. These products represent a shift from general-purpose models to specialized, workflow-integrated solutions designed for specific professional services.

Unsealed court documents from the New York Times' lawsuit against OpenAI and Microsoft reveal that the companies' internal documentation explicitly warned that their data scraping practices were creating a "doom loop"—a self-reinforcing cycle where AI-generated content would degrade web quality, ultimately poisoning the training data that future models depend on. The documents characterize the scraping as "the largest theft of labor" and show awareness of the long-term consequences.

Virginia Governor Abigail Spanberger signed an executive order creating a state AI task force and implementing new restrictions on data center development in Virginia, traditionally the nation's data center hub. The order empowers local communities to have greater influence over large data center projects and slows the approval process for new infrastructure.