AI Daily Newsletter: Anthropic Creates Its Own Medication and Students Receive Lower Test Scores Due to AI Use

🧬 Anthropic Develops Its Own Medicines, Showing AI is Not Just a Chatbot

Anthropic has announced that it will develop its own medicines for diseases that Big Pharma has overlooked due to lack of profitability. At the launch event of Claude Science (an AI workspace for researchers), Anthropic clarified that this timeline aims to support its non-profit mission and help them better understand how AI can be applied in science.

A researcher at UCSF used Claude Science to discover a new virus in just a few minutes - something that his entire team had missed for over a year. Claude also analyzed 100 rare genetic diseases in under an hour and identified 32 candidates for computer analysis.

Source: THE DECODER

🔐 Number of Security Vulnerabilities Increases Dramatically Since AI Started Hunting for Bugs

Epoch AI has released a chart showing that the number of severe security vulnerabilities (CVE) in June 2026 was 3.5 times higher than the previous record. Approximately 1,500 new vulnerabilities were reported by 21 organizations in one month.

This number coincides with the time Anthropic released Claude Mythos Preview in April - the first AI model that can find software vulnerabilities on its own. Anthropic’s “Glasswing” program has discovered over 10,000 severe vulnerabilities. OpenAI also has a similar program called “Daybreak” that is contributing to the increase in this number.

Source: THE DECODER

🏭 Microsoft Invests $2.5B to Establish AI Deployment Company, Sending 6,000 Engineers to Businesses

Microsoft has announced a $2.5 billion investment in a new unit called “Frontier Company” - which will place 6,000 AI engineers to work directly with client businesses. The goal is to integrate AI into measurable business processes, not just experiments.

At the same time, Microsoft is merging its Copilot consumer and enterprise applications into a single app, with an “AutoPilot” agent that automatically handles schedules and emails. Director Jacob Andreou said in an internal memo that the team has “eliminated ineffective things” and Copilot must “earn its right to exist.”

Source: THE DECODER

📚 Study of 26,000 Students: AI Helps with Homework Faster but Reduces Test Scores by 24%

A large-scale study from China with 26,000 high school students shows a two-sided picture of AI in education. After 6 months of using AI, homework scores increased by 18% and completion time decreased from 64 to 45 minutes - but test scores decreased by up to 20%.

More seriously, the negative impact on important entrance exams took 2 years to fully manifest, with a decrease of 18-24%. The percentage of students self-reporting AI use increased from nearly 0% to 80% during the study period. The most popular tools were Doubao (ByteDance), DeepSeek, ChatGLM (Baidu), and Qwen (Alibaba).

Source: THE DECODER

🔬 UK AI Security Institute: Current Benchmarks Underestimate AI Agent Capabilities

The UK AI Security Institute (AISI) has announced an important finding: current standard benchmarks are systematically underestimating the capabilities of AI agents. The reason is that these benchmarks set computational limits (token budget) too low, not reflecting the true abilities.

When increasing the token budget by 10 times, the success rate in software engineering tasks increased by about 25%. Newer models benefit more from increased compute budget. According to AISI, the actual development speed of AI agents is about 60% faster than what previous measurements showed.

Source: THE DECODER