Daily AI Builder Brief - June 4, 2026

Executive Summary

Claude Code 2.1.162 lands waitingFor in claude agents --json and gives explicit WebFetch(domain:...) rules precedence over preapproved hosts.
2.1.162 also fixes MCP timeout values below 1000ms being floored, so sub-second timeouts now stick.
2.1.160 prompts before writing build-tool config (.npmrc, bunfig.toml, .bazelrc) and shell startup files. Your agents can't silently rewrite tooling anymore.
Anthropic's year-long cyber-threat study: AI adoption among medium/high-risk actors jumped from 33% to 56%, with 67% using AI for malware development.
Project Glasswing expanded to ~150 orgs and added Claude Security for codebase scans and patch suggestions; partners have found 10,000+ high/critical flaws.
Heads up on billing: programmatic Claude usage on subscription plans moves to a separate monthly credit pool starting June 15.

Claude Code

2.1.162: agent visibility and WebFetch precedence

claude agents --json now reports what a waiting session is blocked on via waitingFor. Explicit WebFetch(domain:...) permissions now override preapproved hosts, and the sub-1000ms MCP timeout-flooring bug is fixed. changelog

2.1.161: parallel tool resilience

A failed Bash command no longer cancels other calls in the same parallel batch. /mcp now hides unused claude.ai connectors behind "Show unused connectors." changelog

2.1.160: config-write guardrails

acceptEdits now prompts before writing .npmrc, .yarnrc*, bunfig.toml, .bazelrc, and pre-commit config. Single-file grep now satisfies the read-before-edit check. changelog

Anthropic

Mapping a year of AI-enabled cyber threats

832 banned accounts, 13,873 actions, 482 techniques mapped to MITRE ATT&CK. Conclusion: attacks are getting more autonomous and ATT&CK no longer captures AI-enabled methods. Shared with Verizon's 2026 DBIR. report

Project Glasswing expansion plus Claude Security

Mythos Preview now reaches ~150 orgs; Claude Security adds automated codebase scans and patch suggestions. news

MCP Ecosystem

2026-07-28 spec release candidate: validation window open

No net-new MCP news in 48h, but the RC validation window is live. Stateless core, MCP Apps (server-rendered UIs), a Tasks extension for long-running work, OAuth/OIDC-aligned auth, and a formal deprecation policy. Tier 1 SDKs ship support before the July 28 final. Test now. spec RC

Broader AI

Nothing today.

What This Means For Builders

Audit your acceptEdits automations: 2.1.160 will now interrupt agents that touch tooling config. Pre-approve only what you trust.
If you run remote MCP servers, prototype against the stateless RC now. Sticky sessions become optional, which simplifies your load balancing.
The 33% to 56% attacker-adoption jump means your threat model should assume AI-accelerated malware against your own deploys. Treat agent permissions as attack surface.
Plan for June 15: move programmatic Claude usage off shared subscription quota before the credit-pool split bites.