deepseek unveils new tools as anthropic ships new models

Two new studies are raising serious concerns about the reliability of AI chatbots for health advice. Researchers found that roughly half of chatbot responses to health questions were problematic, with nearly 20 percent of answers deemed highly problematic and potentially harmful. Grok, trained partly on social media content from X, was the most likely to give dangerous responses. In a separate test, AI models failed over 80 percent of the time at listing possible diagnoses like a doctor would. Experts advise people to share chatbot answers with their doctors rather than relying on them alone.

On the enterprise front, Shiven Ramji, president of Okta's customer identity cloud business, warns that shadow AI agents are creating hidden risks for organizations. Many teams experiment with AI without IT oversight, leaving companies blind to where agents are, what they can connect to, and what they can do. Without clear answers, organizations risk over-provisioned access and uncontrolled automation. Meanwhile, two recent vulnerability disclosures in Claude and OpenClaw show that a single manipulated input can silently compromise everything an AI agent can reach. Nearly 74 percent of companies plan to deploy agentic AI within two years, but governance has not kept pace.

In other AI developments, DeepSeek has released its new V4 Pro and V4 Flash models, both open-source. The company says V4 Pro rivals top closed-source models in reasoning and trails only Gemini-3.1-Pro in rich world knowledge. V4 Flash offers faster response times while closely approaching V4 Pro in reasoning. Sakana AI is accepting applications for the beta test of Fugu, its multi-agent orchestration system that coordinates frontier models for coding, math, and scientific reasoning tasks. The system is available as an API with two versions: Fugu Mini for low latency and Fugu Ultra for maximum performance.

AI is also entering a new phase in healthcare, where doctors can now code custom clinical workflow tools using agentic AI like Claude Code. This shift toward doctor-led software development requires new security audits and professional engineering oversight. At the same time, AI-driven vulnerability discovery is reshaping cybersecurity, with Anthropic's Claude Mythos able to detect system vulnerabilities. Healthcare must prepare for a new era of AI-enabled cyberattacks. In a lighter note, Sony's Aibo robot defeated a professional table tennis player 3-0 at a Tokyo event, with Dr. Hiroshi Ishiguro calling it a major breakthrough in AI technology.

The Vatican is moving faster than most legacy institutions to shape rules and guardrails for verifying reality in the age of AI. Church leaders are increasingly warning about the challenges posed by artificial intelligence amid unusual geopolitical and digital clashes. Meanwhile, the Grossmont Union High School District in California is seeking AI training for teachers, highlighting the growing need for AI education in schools. The School District of Greenville County in South Carolina wants a platform to track instructional spending, and Belvidere Community Unit School District 100 in Illinois is looking for a multilingual program review agency. Deadlines for these solicitations range from May 5 to May 12, 2026.

Key Takeaways

  • Two studies found that about half of AI chatbot health responses are problematic, with nearly 20 percent potentially harmful.
  • Grok was the most likely to give highly problematic health answers, possibly due to training on social media content from X.
  • AI models failed over 80 percent of the time at listing possible diagnoses like a doctor would.
  • Shiven Ramji warns that shadow AI agents operating outside IT oversight create hidden risks for enterprises.
  • Two vulnerability disclosures in Claude and OpenClaw show that a single manipulated input can compromise everything an AI agent can reach.
  • Nearly 74 percent of companies plan to deploy agentic AI within two years, but governance has not kept pace.
  • DeepSeek released open-source V4 Pro and V4 Flash models, with V4 Pro rivaling top closed-source models and trailing only Gemini-3.1-Pro in world knowledge.
  • Sakana AI opened beta applications for Fugu, a multi-agent orchestration system available as an API with Fugu Mini and Fugu Ultra versions.
  • Sony's Aibo robot defeated a professional table tennis player 3-0 at a Tokyo event.
  • The Vatican is moving to set rules for AI, and California's Grossmont Union High School District is seeking AI training for teachers.

Studies warn AI health answers often wrong

Two new studies show that AI chatbots give problematic health answers about half the time. One study found nearly 20 percent of responses were highly problematic and could cause harm. Another study showed AI models failed over 80 percent of the time at listing possible diagnoses like a doctor would. Researchers advise people to share chatbot answers with their doctors rather than relying on them alone.

Researchers urge caution with AI health questions

Researchers found that Grok was most likely to give highly problematic health answers, possibly because it is trained partly on social media content from X. One study showed about half of chatbot responses to health questions were problematic. Another study found an 80 percent failure rate when AI models tried to list possible conditions for a patient's symptoms like a doctor would.

Sakana AI opens beta for Fugu multi-agent system

Sakana AI is accepting applications for the beta test of Sakana Fugu, its multi-agent orchestration system. Fugu coordinates pools of frontier foundation models for coding, math, and scientific reasoning tasks. It is available as an API with two versions: Fugu Mini for low latency and Fugu Ultra for maximum performance. The system learns to dynamically assemble agents and coordinate them through efficient collaboration patterns.

Vatican moves to set rules for AI

The Holy See is moving faster than most legacy institutions to shape rules and guardrails for verifying reality in the age of AI. The Vatican has stepped up efforts and implemented formal measures. Church leaders are increasingly warning about the challenges posed by artificial intelligence amid unusual geopolitical and digital clashes.

Shadow AI agents create hidden risks for enterprises

Shiven Ramji, president of Okta's customer identity cloud business, warns that many organizations lack visibility into shadow AI agents operating outside IT oversight. Teams independently experiment with AI without understanding security implications. Ramji says CIOs need to answer three questions: where are the agents, what can they connect to, and what can they do. Without clear answers, organizations risk over-provisioned access and uncontrolled automation.

Sony AI robot beats professional table tennis player

Sony Corp. announced that its robot, the Sony Aibo, defeated a professional table tennis player in a match. The robot used artificial intelligence to track and hit the ball and won the match 3-0. The event was held at a Sony event in Tokyo with about 100 people in attendance. Dr. Hiroshi Ishiguro from Osaka University called it a major breakthrough in AI technology.

DeepSeek releases new AI model with world-class reasoning

DeepSeek has released its new V4 Pro and V4 Flash AI models, both open-source. DeepSeek says V4 Pro has enhanced agentic capabilities and rivals top closed-source models in reasoning. V4 Flash has faster response times and closely approaches V4 Pro in reasoning abilities. The company claims V4 Pro trails only Gemini-3.1-Pro in rich world knowledge.

AI enters new phase in healthcare on two fronts

Artificial intelligence is becoming user-friendly enough that doctors can code custom clinical workflow tools using agentic AI like Claude Code. This shift toward doctor-led software development requires new security audits and professional engineering oversight. Meanwhile, AI-driven vulnerability discovery is reshaping cybersecurity, with Anthropic's Claude Mythos able to detect system vulnerabilities. Healthcare must prepare for a new era of AI-enabled cyberattacks.

Claude and OpenClaw flaws reveal AI agent risks

Two recent vulnerability disclosures in Claude and OpenClaw show that a single manipulated input can silently compromise everything an AI agent can reach. The Oasis Threat Research Team uncovered attack chains called Claudy Day for Claude and a localhost trust exploit for OpenClaw. Both attacks worked because of design assumptions that allowed attackers to take control without the user noticing. Nearly 74 percent of companies plan to deploy agentic AI within two years, but governance has not kept pace with deployment.

California district seeks AI training for teachers

Grossmont Union High School District in California is looking for professional development focused on artificial intelligence. The School District of Greenville County in South Carolina wants a platform to track and manage instructional spending. Belvidere Community Unit School District 100 in Illinois is searching for an agency to conduct a multilingual program review. The deadlines for these solicitations range from May 5 to May 12, 2026.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

AI Healthcare Chatbots Artificial Intelligence Machine Learning Deep Learning Natural Language Processing Robotics Table Tennis Sony Aibo Shadow AI Security Cybersecurity Vulnerability Discovery Anthropic Claude OpenClaw Oasis Threat Research Team Agentic AI Governance Education Teacher Training California Grossmont Union High School District

Comments

Loading...