AI agents were at the center of this week’s developments, opening new possibilities in business and research while putting pressure on existing security practices.

That security challenge was underscored by attacks involving the ARTEX agentic penetration-testing framework, which was used in a ConnectWise supply-chain compromise affecting more than 2,250 Microsoft 365 tenants. Attackers also exploited two PHP vulnerabilities rated 9.2 to access vulnerable web servers, while researchers reported AI agents extracting password hashes and MFA secrets from a European government system in just two hours.

The speed of such attacks is straining conventional defenses and prompting calls for automated detection, tighter access controls and network segmentation. Researchers also identified weaknesses in an earlier version of OpenAI’s internal agent-monitoring system; OpenAI says it patched production safeguards, though researchers have not independently evaluated the changes.

Companies are also trying to make agents safer to use: Anthropic is making Claude’s auto mode the default for Pro, Max and Team users, with automated checks designed to approve routine actions and block risky ones. Anthropic has contributed its Agent Skills format to the Linux Foundation’s Agentic AI Foundation; more than 70 platforms have adopted it, and over a million skills have been published.

Access policies are diverging across commerce and finance. Walmart’s Agent Gateway lets approved agents shop within customer-set spending and approval limits, while Amazon and several banks have restricted agent access; retailers are also weighing how to protect customer relationships as shopping shifts toward agents.

The broader debate over AI oversight is intensifying: Anthropic proposed federal AI safety legislation, while Google DeepMind employees resigned over the company’s military AI work.

In life sciences, OpenAI announced plans to acquire Orchestra Bio and integrate its drug-development workflow platform with Rosalind, its research tools, to help scientists coordinate experiments and collaboration in a process that typically takes 10 to 15 years.

Anthropic reported a protein-design experiment in which a model generated 318,484 candidate designs targeting scorpion venom in 64 hours. Lab tests of 211 designs found 17 that bound to toxins, but researchers have not established whether those designs can neutralize venom.

AI’s potential in medicine also drew attention: a study found Microsoft’s MAI Diagnostic Orchestrator correctly diagnosed 83.2% of difficult cases, versus 36.1% for generalist physicians. The comparison did not include specialists, and the system still needs clinical validation.

The growth of AI is also driving large data-center investments while sharpening electricity constraints. NVIDIA has made or guaranteed $164.5 billion in commitments supporting AI data centers, and projects in Ohio, Kentucky and elsewhere are advancing; a proposed Essex data center, however, faces grid delays extending into the 2030s.