AI news5
FERC issues six unanimous orders mandating fast-track grid access for AI data centers
The Federal Energy Regulatory Commission approved six unanimous orders on June 18 requiring the six major US grid operators (PJM, MISO, SPP, CAISO, ISO-NE, NYISO) to eliminate interconnection bottlenecks for AI data centers. The orders cover grid infrastructure serving 200 million Americans across 30+ states and the District of Columbia, set 30-day deadlines for operator reliability reports, and impose 60-day deadlines for pricing reform. Critically, FERC has formally barred grid operators from spreading AI data center connection costs to ordinary ratepayers, with federal enforcement authority behind the requirement. The orders also extend a co-location framework, allowing data centers to connect directly behind the meter at power generation facilities, bypassing the congested interconnection queue entirely.
ferc.gov ↗Why it matters
Six simultaneous federal mandates targeting the six major US grid operators is the most direct intervention Washington has made in AI infrastructure to date, and the ratepayer protection provision removes a political obstacle that has stalled similar reforms in state legislatures.
MosaicLeaks: deep research AI agents leak private data in 34% of responses
ServiceNow and HuggingFace researchers published MosaicLeaks, a study showing that deep research agents combining private documents with web retrieval leak sensitive information in 34% of responses. Their proposed fix, Privacy-Aware Deep Research (PA-DR), uses reward-based safe query construction instead of user prompting to reduce leakage from 34% to 9.9% while maintaining task success rates. The reward-based approach is more robust than instruction-based privacy enforcement because it does not depend on users correctly phrasing privacy constraints in their prompts.
huggingface.co ↗Why it matters
Most enterprise AI deployments combine internal documents with external retrieval, making the 34% default leakage rate an immediate audit priority for any organization running research-style agents on sensitive data.
Google DeepMind publishes AI Control Roadmap for securing enterprise AI agents
Google DeepMind released a framework describing how organizations should build and manage AI agent deployments securely, covering sandboxing, endpoint security, prompt injection resistance, and cross-agent communication monitoring. The roadmap treats internal agents as potentially misaligned by default rather than trusting alignment as a sufficient safeguard, providing system-level assurance independent of whether individual agents behave as intended. Key recommendations include isolating every agent action surface and applying endpoint security controls to all tool calls.
deepmind.google ↗Why it matters
A security-first framework from the team running some of the world's largest AI deployments sets a credible baseline for enterprise organizations building agentic systems, and treating misalignment as the default threat model is a meaningful shift from prior approaches.
Amazon Bedrock AgentCore Harness reaches general availability
AWS launched Amazon Bedrock AgentCore Harness as generally available on June 19, providing a managed production runtime for enterprise AI agents with built-in identity, memory, guardrails, and an MCP-compatible web search connector. The web search tool lets agents query the open web securely from within AWS without managing external API credentials. The launch targets engineering teams that want to move from prototype to production agent without custom-building the operational scaffolding for each component.
aws.amazon.com ↗Why it matters
Managed agent runtimes from AWS shift the operational scaffolding of enterprise AI deployment into cloud-provider territory, which compresses the build time for production agents but also concentrates vendor dependency in a single platform.
Anthropic adds artifacts to Claude Code: live shareable visual pages from coding sessions
Claude Code's new artifacts feature turns active coding sessions into live, shareable visual pages suited for PR walkthroughs, architecture diagrams, and system explainers. Artifacts refresh automatically as the underlying session updates, maintain version history, and include privacy controls for managing who can access the page. The feature is currently in beta for Team and Enterprise plan subscribers.
claude.com ↗Why it matters
Shareable, auto-updating session artifacts reduce the friction between agentic coding work and team communication, making it practical for engineering teams to build async review processes around Claude Code outputs without exporting static documents.
AI in the nonprofit sector4
HAIDI Innovation Track offers up to KES 5 million for African AI-disability inclusion projects
The Hub for Artificial Intelligence and Disability Inclusion, part of the AI4D multi-funder development programme, is accepting applications through July 17, 2026 for African organizations with post-MVP AI solutions designed for and with disability communities. Selected participants receive up to KES 5 million (approximately $38,000 USD), facilitated user testing with disability communities across Africa, access to the AI4D international funding network, and structured technical and strategic support. The programme explicitly requires that solutions be in active use rather than prototype stage.
nerdzfactory.org ↗Why it matters
Disability-focused AI receives a fraction of global AI investment, and a multi-funder programme designed specifically for African development contexts and community co-design requirements is uncommon; organizations building in this space have few other structured pathways to funding and validation.
Special Olympics study finds 5 major AI systems consistently portray people with intellectual disabilities as dependent and childlike
A peer-reviewed study published June 16 in the Disability and Health Journal by Special Olympics and Oregon State University analyzed 25,000 AI-generated stories about people with intellectual disabilities, using GPT-4-Turbo to score representation patterns. All five tested AI systems consistently generated portrayals showing greater dependence, childlike characteristics, and need for supervision compared to portrayals of people without disabilities. The researchers describe the pattern as "automated ableism" built into the training cycle, and the study is the first peer-reviewed research led and published by Special Olympics.
specialolympics.org ↗Why it matters
AI systems are increasingly used in healthcare intake, social services assessments, and educational settings; systematic ableist bias in underlying models will shape consequential decisions affecting one of the most marginalized populations unless explicitly audited and corrected upstream.
GLAAD releases "Build for Everyone," its first comprehensive AI impact report on LGBTQ communities
GLAAD published its inaugural AI report on June 17, documenting how AI systems affect LGBTQ people across healthcare access, hiring decisions, content moderation, and online visibility. A case study documents AI models surfacing misinformation about conversion "therapy" when users ask healthcare questions. The report uses GLAAD's 40-year methodology for measuring LGBTQ representation in media as the foundation for an AI audit framework, and concludes with graded recommendations for AI developers and deployers.
glaad.org ↗Why it matters
GLAAD's entrance into AI accountability auditing brings one of the sector's most established nonprofit research frameworks directly into AI governance, and its findings on healthcare misinformation and content moderation bias have immediate policy implications for LGBTQ-serving organizations navigating AI tool adoption.
Telepatia raises $33 million Series A from a16z for AI clinical platform across Latin America
Andreessen Horowitz led a $33 million Series A in Telepatia, an AI clinical platform deployed across 25-plus hospital systems in Brazil, Colombia, and Mexico, reaching 14 million patients since its July 2025 launch. Protocol adherence among deploying physicians has risen from 84% to 99%, physicians recover an average of 1.7 hours per day, and Telepatia reports 60,000 medical errors prevented in real time. The platform is designed around Latin America's structural clinician shortage: the region has 43% fewer physicians per capita than OECD countries, and Telepatia was built for that constraint from the ground up rather than adapted from a US-built system.
a16z.com ↗Why it matters
AI clinical platforms built natively for under-resourced health systems represent the category most likely to achieve durable impact on global health equity, and a16z's lead position signals significant follow-on capital; the real-time error prevention data also provides one of the strongest efficacy cases in nonprofit-adjacent health AI to date.