The evening briefing.
Today across AI and tech: Anthropic enters drug discovery, ByteDance reveals new scaling laws, and Microsoft pushes its Copilot super app.
Frontier AI Applications. Anthropic is expanding into drug discovery with Claude Science, an AI workbench for scientists. This initiative aims to accelerate scientific discovery and healthcare interventions, leveraging AI models for complex research tasks. Meanwhile, Google DeepMind has partnered with A24 for a research collaboration, exploring AI's role in creative industries. These moves highlight a growing trend of major AI labs applying their advanced models to new, high-impact domains beyond traditional software.
AI Capabilities & Evaluation. ByteDance researchers have uncovered a new scaling law, suggesting AI agents can double their learning speed every three months by interacting with real-world tasks. This discovery could extend the current AI boom as traditional development methods face limitations. However, the UK's AI Security Institute warns that standard benchmarks often underestimate actual AI agent capabilities. Their study found success rates jumped significantly when compute budgets were increased, indicating current evaluations may not fully capture frontier progress.
Geopolitics & Security. Geopolitical tensions are impacting AI development, as Anthropic faces a complicated China problem with Claude Code. The company attempts to block Chinese firms, while Alibaba has banned its own employees from using the tool due to hidden code concerns. Concurrently, a sharp rise in security vulnerability reports has been observed, with Epoch AI reporting a 3.5x spike in CVEs since AI models began actively hunting for bugs. This highlights both the power and potential risks of AI in cybersecurity.
Societal & Economic Impact. The UK's National Crime Agency and Internet Watch Foundation warn parents against publicly sharing children's images online due to rising AI-generated sexual abuse fears. On the economic front, AI's impact on white-collar jobs is becoming clearer, with one developer noting a significant drop in course sales due to AI uncertainty. Meanwhile, SAP is reallocating budgets to create new AI-powered roles, signaling a shift in corporate strategy towards AI integration and new job creation.
Anthropic aims to develop its own drugs with Claude Science
Anthropic has launched Claude Science, an AI workbench designed for scientists to accelerate drug discovery and healthcare interventions. The new platform integrates fragmented tools and datasets, generating figures and visuals to streamline research.
ByteDance discovers new scaling law for AI agent improvement
Researchers at ByteDance's Seed AI team have found that AI agents can double their learning speed every three months through real-world task interactions. This new scaling law could help sustain the AI boom as traditional development methods reach their limits.
Current AI launches Open Source AI Gap Map
Current AI, a non-profit backed by $400 million, has launched its Gap Map v0.1, an index detailing 421 open-source AI products. The map covers 266 software tools, 85 models, 50 datasets, and 20 hardware projects from 228 organizations.
Microsoft enters AI super app race with overhauled Copilot and AutoPilot agents
Microsoft reportedly plans to merge its consumer and enterprise Copilot apps into a single offering by August, cutting rarely used features. New AI agents called "AutoPilot" will handle background tasks for an additional fee.
Claude Code faces China problem amid bans and hidden code concerns
Anthropic is attempting to block Chinese companies from accessing Claude Code, but firms are bypassing restrictions via VPNs and overseas subsidiaries. Alibaba has also banned its employees from using the tool after discovering hidden code that could identify Chinese users.
UK parents warned over sharing children's images due to AI sexual abuse fears
The National Crime Agency and Internet Watch Foundation have issued guidance advising UK parents not to post photos of their children online. This warning comes amid a rising threat of AI-generated sexual abuse material.
Security vulnerability reports explode with AI bug hunting models
Epoch AI reports a significant surge in security vulnerability reports, with 21 organizations reporting about 1,500 high-severity and critical CVEs in June 2026. This increase, 3.5 times the previous monthly record, coincides with the launch of AI-powered bug-hunting programs.
UK's AI Security Institute finds benchmarks underestimate AI agent capabilities
A study by the UK's AI Security Institute reveals that standard AI evaluations systematically underestimate agent capabilities by capping compute budgets. Success rates on software engineering tasks jumped approximately 25 percent when the token budget was increased tenfold.
Apple's Safari now allows control by AI agents
Apple's WebKit team has shipped Safari Technology Preview 247 with a built-in Model Context Protocol server. This integration enables AI agents to control Safari, marking a significant step for on-device AI capabilities within the browser.
Google DeepMind and A24 announce research partnership
Google DeepMind and A24 have announced a first-of-its-kind research partnership. The collaboration aims to explore new frontiers in AI research, particularly within creative industries.
Quantum Systems raises $1.2B, IQM lists on major US exchange
Quantum Systems has raised $1.2 billion in funding, while IQM has become the first European quantum company to list on a major US exchange. These developments highlight significant investment and growth in the quantum computing sector.
SAP shifts focus to AI-powered jobs, cuts other budgets
SAP is reallocating its budget and hiring efforts to focus on creating new AI-powered jobs, while slashing travel and expenses. This strategic shift aims to increase AI spend amid broader concerns about software longevity.
AI impacts developer course sales, raising job market concerns
A developer reported that sales for their programming courses are down significantly, attributing the decline to AI. Many people are questioning the future of developer jobs and are reluctant to invest in new skills.
Contrastive Decoding Diffing recovers finetuning data from LLM logits
Researchers developed Contrastive Decoding Diffing (CDD), a method that recovers verbatim finetuning data from narrowly finetuned LLMs using only grey-box logit access. This technique does not require access to model weights or activations.