Daily AI News

Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems

Mistral AI unveiled Leanstral 1.5, a 119 billion‑parameter mixture‑of‑experts code‑agent for the Lean 4 theorem‑proving language, released under the Apache‑2.0 license. The model activates 6.5 billion parameters per token and solves 587 of 672 PutnamBench problems, outperforming prior open models on miniF2F benchmarks.

Read source →

Anthropic wants to develop its own drugs

Anthropic announced Claude Science, an AI‑powered workbench that stitches together disparate scientific tools and datasets, automatically generates figures and visualizations, and is targeted at accelerating drug‑discovery workflows.

Read source →

Microsoft follows Anthropic and OpenAI into the AI super app race with overhauled Copilot and AutoPilot agents

Microsoft plans to combine its consumer and enterprise Copilot experiences into a single super‑app in August, sunsetting low‑usage features and introducing AutoPilot agents that execute background tasks for a fee.

Read source →

Takeda signs US$600M AI drug discovery deal with Insilico

Japanese pharmaceutical giant Takeda entered a $600 million partnership with Hong‑Kong AI firm Insilico Medicine to apply generative AI to early‑stage drug discovery across multiple therapeutic areas.

Read source →

Chinese AI video maker Kling raises $2 billion as it gears up for Hong Kong IPO

Kuaishou’s generative‑video division Kling secured roughly $2 billion in financing ahead of a planned Hong Kong IPO, aiming to scale its AI‑powered video creation platform globally.

Read source →

Security vulnerability reports have exploded since AI models started hunting for bugs

Epoch AI reported a 3.5‑fold surge in high‑severity CVE disclosures in June 2026, with 21 organizations filing about 1,500 critical reports, a spike linked to the deployment of AI‑driven bug‑hunting programs.

Read source →

UK's AI Security Institute finds standard benchmarks systematically underestimate what AI agents can actually do

A study of seven common AI benchmarks by the UK AI Security Institute showed that limiting token budgets suppresses agent performance, with software‑engineering success rates climbing ~25 % when compute budgets were increased tenfold.

Read source →

GPT and Claude failed Bridgewater's finance tests because the right answers were never public

Bridgewater and Thinking Machines Lab fine‑tuned the Qwen3‑235B model for finance, attaining 84.7 % accuracy—outperforming Gemini, Claude, and GPT—while costing a fraction of the alternatives; the test also revealed that many benchmark answers are not publicly disclosed.

Read source →

Google DeepMind Unionization Talks Are Off to a Rocky Start

Employees at Google DeepMind reported a turbulent start to union negotiations, citing executive resistance and uncertainty about labor rights within the elite AI research lab.

Read source →

Meta's AI agent push is moving slower than Zuckerberg planned

In an internal town hall, Meta’s leadership admitted that its AI‑agent initiatives are lagging behind internal timelines, despite public optimism, prompting concerns about execution speed relative to competitors.

Read source →