AI News: April 11, 2026

1. CIA Produces First Fully Autonomous Intelligence Report, Plans AI “Coworkers” Across All Platforms

Government. CIA Deputy Director Michael Ellis revealed the agency recently produced its first fully autonomous intelligence report using AI. Within the next two years, the CIA plans to embed AI coworkers into all analytic platforms. Within a decade, officers will manage teams of AI agents as “autonomous mission partners.” Ellis cautioned the CIA “cannot allow the whims of a single company” to constrain its AI use, an indirect criticism of Anthropic’s contractual restrictions on lethal applications. Source

2. LLMs Excel at Coding and Math but Struggle with Casual Questions, and Reinforcement Learning Explains Why

Research. A new analysis explains why LLMs dominate coding and math benchmarks but stumble on everyday questions. Reinforcement learning with verifiable rewards (RLVR) only works well where outputs can be automatically verified (code compiles, math checks out). Fuzzy domains like writing or consulting lack clean reward signals, revealing a fundamental architectural limit in current training approaches rather than a fixable oversight. Source

3. Leaked “SteamGPT” Files Reveal Valve Building Internal AI Tool for Support and Anti-Cheat

Gaming. Data miner Gabe Follower discovered code references to “SteamGPT” in a Steam VR Beta update released April 7. The internal AI tool appears designed to automate customer support ticket handling and connect to Valve’s Trust Score system for CS2 anti-cheat behavior analysis. Code snippets reference Trust_GetTrustScoreInternal, CSbot, and player_evaluation. References were quickly removed after discovery. Source

4. Critical Marimo Python Notebook Vulnerability Exploited Within 10 Hours of Disclosure

Security. A critical pre-auth remote code execution vulnerability (CVE-2026-39987, CVSS 9.3) in open-source Python notebook tool Marimo was exploited in the wild within 9 hours and 41 minutes of disclosure. The terminal WebSocket endpoint lacks authentication, allowing unauthenticated attackers to obtain a full PTY shell. Particularly concerning because notebook tools often sit close to model-building and analytics infrastructure. Fixed in version 0.23.0. Source

5. AI Therapy Chatbot Bans Accelerate: Maine and Missouri Move on Legislation

Policy. Maine lawmakers approved LD 2082 prohibiting clinical use of AI in mental health therapy, heading to the governor with the legislature adjourning April 15. Missouri’s HB 2372 includes a therapy chatbot ban with a $10,000 first-violation penalty, approved by the full House and now in Senate committee. Both follow Tennessee’s signed SB 1580, signaling a national trend of states restricting AI from acting as mental health professionals. Source

6. UC San Diego Chip Design Could Slash Data Center Power Waste for GPU Workloads

Hardware. UC San Diego researchers published a prototype chip design in Nature Communications that achieves 96.2% efficiency converting 48V to 4.8V for GPUs using piezoelectric resonators combined with capacitors. The approach delivers four times more current than previous piezoelectric designs and could lead to smaller, more energy-dense power converters for AI data centers, addressing a critical bottleneck in infrastructure expansion. Source

7. Ex-DeepMind Researchers Launch Elorian Visual AI Startup with $55M

Venture Capital. Former Google DeepMind researchers including Andrew Dai debuted Elorian, a Palo Alto-based visual AI startup that raised $55M at a $300M valuation. The company focuses on enabling AI systems to understand and reason through visual data including spatial relationships, with applications in architecture, automotive, and robotics. Source