AI Daily Newsletter: Claude Code Refutes Doctors and GLM 5.2 Surpasses Claude in Security

🔥 1. GLM 5.2 surpasses Claude for the first time in security benchmark

Zhipu AI launched GLM 5.2 — a 750B parameter MoE model (40B active), MIT license, 1M token context — and it defeated Claude Code by 7 F1 points in the IDOR vulnerability detection task on Semgrep’s benchmark. The cost is only ~$0.17 per vulnerability found, 6 times cheaper than frontier models. This is the first time an open-weight model on a prompt surpasses a commercial coding agent in a security reasoning-heavy task.

Source: Semgrep Blog

🤖 2. Claude Code reads MRI, contradicts doctor’s diagnosis

A programmer used Claude Code (Opus 4.8) to analyze 266MB of raw DICOM data from an MRI scan of the shoulder — and the AI concluded “no tendon tear” while the doctor diagnosed a Grade III tear. The arbitration result leaned towards the AI. The story raises a disturbing question: if AI can read MRI scans better than doctors in some cases, where should patients place their trust?

Source: antoine.fi

⚡ 3. Brown University: 50+ students cheat with AI, grades plummet

Professor Roberto Serrano discovered that at least 50 students cheated on the ECON 1170 exam. The take-home midterm had an average score of 96/100, with 40 perfect scores. The in-person final exam had an average score of 48/100, with 22 students who had previously scored 100 on the midterm not even showing up. Brown remains silent, while Princeton has abolished its 133-year-old honor code for similar reasons.

Source: EL PAÍS

🛡️ 4. China puts supercomputer LineShine at #1 on TOP500

For the first time in 9 years, China has submitted a system to the TOP500 list — and it’s ranked #1. LineShine achieved 2.198 Exaflops FP64 (Rmax) with over 22,000 nodes, 13 million CPU cores, all using domestic Armv9 LX2 CPUs with 304 cores. It also led the HPCG benchmark. The system consumes 42.22 MW — less efficient than its competitors but enough to put pressure on the US DOE budget.

Source: Chips and Cheese

📊 5. OpenAI launches its first chip Jalapeño + Trump admin releases Anthropic Mythos

OpenAI introduced its first custom chip “Jalapeño” manufactured by Broadcom — marking a shift away from Nvidia. At the same time, the Trump administration released Anthropic Mythos to over 100 US companies and agencies, while the export ban on Anthropic to Asia continues — prompting a wave of Asian AI startups to build similar models.

Source: TechCrunch AI