Security pros ask the US to lift the Fable and Mythos ban, and the week in Claude
PLUS: Anthropic pauses its Agent SDK billing change, Claude Code can now publish live pages in beta, and Claude Design now builds from your design system
Hey, friends. Here's your week in Claude, from Jun 14 to Jun 20.
This week, security experts asked the US to lift the Fable ban. Anthropic paused its Agent SDK billing change. And Claude Code can now publish live pages from your session, in beta.
Plus 10 quick hits. Every claim has a link.
Security pros ask the US to lift the Fable and Mythos ban
Dozens of security experts signed an open letter, per TechCrunch. It asks the US to lift its export controls on Fable 5 and Mythos 5.
The letter is dated Jun 14. It's addressed to Secretary Lutnick and National Cyber Director Cairncross. It says the action "has taken the best models away from defenders."
It also asks for an open, scientific and transparent process for AI risk assessments. Per TechCrunch, 76 experts had signed as of Jun 15. They include Alex Stamos, Katie Moussouris and Rachel Tobac.
On Jun 12, Anthropic said a US export control directive covers access by any foreign national, in or outside the US. It said the net effect was that it had to disable both models for all customers. Its other models weren't affected.
Anthropic said it was complying. It disagreed that a narrow potential jailbreak should be cause for recalling the model.
Anthropic has said the order may stem from a reported jailbreak of Fable.
Per TechCrunch, Moussouris says Amazon researchers showed the method in a paper. It isn't public, but she has reviewed it. In a blog post, she wrote that it didn't show a real jailbreak.
Anthropic pauses its Agent SDK billing change
A planned change to Agent SDK billing on paid plans didn't happen on Jun 15.
That day, Anthropic's help article said it had "paused the previously-announced changes." For now, Agent SDK, claude -p and third-party app usage still draw from your plan's limits.
The change would have moved that usage to a separate monthly credit, per eWeek (May 18). Pro would have gotten $20, Max 5x $100 and Max 20x $200.
Claude Code can now publish live pages from your session, in beta
Claude Code got live pages, finer permission rules and a faster way to change settings.
It shipped versions 2.1.178 to 2.1.183 from Jun 15 to 19, per its weekly digest.
- Artifacts are live pages built from your session, like a PR walkthrough or a dashboard. They update in place as the session works, per Anthropic's Jun 18 post.
- Artifacts are in beta on Team and Enterprise plans, from the CLI and desktop app. Each is private to its author by default. Only signed-in members of your org can view one, and none can be made public.
- Permission rules for deny and ask can now match a tool's inputs. For example, Agent(model:opus) blocks Opus subagents.
- /config key=value sets any setting from the prompt, such as /config thinking=false.
- Auto mode now blocks git reset --hard and similar commands unless you asked to discard local work. It blocks terraform destroy unless you asked for that stack.
Claude Design now builds from your design system
Anthropic updated Claude Design on Jun 17. It now sticks to your design system. It also shares limits with the rest of Claude.
Per Anthropic's post, you can bring in one or several design systems. They can come from a GitHub repo, design files or uploads. Claude builds with your parts and checks its work against your system.
For larger teams, a new admin role can approve one standard system and lock edits. A new editor lets you drag, resize and align. You can hand a design to Claude Code, which picks up from your work.
Claude Design now shares usage limits with your other Claude work, including Claude Code. Anthropic says most people will get much more headroom. Over one million people used Claude Design in its first week, per Anthropic.
Quick hits
- Claude Sonnet 4 and Claude Opus 4 are retired on the Claude API (Jun 15). Requests to them now return an error, per the release notes. Researchers can ask for ongoing access.
- Seven Claude SDKs now support the code execution tool version with REPL state persistence (Jun 18). It needs no beta header, per the release notes. It works on Fable 5, Mythos 5, and Opus and Sonnet 4.5 and newer.
- Workload Identity Federation is generally available on the Claude Platform (Jun 17). It swaps static API keys for short-lived, scoped credentials. It works with any OIDC-compliant identity provider, and API keys still work too.
- Admins can now set up MCP connectors for a whole org through Okta (Jun 18). Users get access on first login, per Anthropic. It's in beta for Team and Enterprise plans, and you apply for access.
- Anthropic opened its Seoul office (Jun 17). It signed an AI safety MOU with Korea's Ministry of Science and ICT. Per Anthropic, NAVER has deployed Claude Code across its entire engineering organization.
- Anthropic joined Frontier, the carbon removal group (Jun 17). Frontier's buyers committed $915 million more, for a $1.8 billion total. TechCrunch reports Anthropic is the first AI startup to join.
- Nobel laureate John Jumper is joining Anthropic from Google DeepMind (Jun 19), per TechCrunch. He led the AlphaFold team. In his post on X, he says he'll take some time to recharge first.
- Claude Code now warns when your requested model is deprecated (2.1.183, Jun 19). It also warns when the model was automatically updated to a newer one.
- Claude Code's auto mode now checks subagents before they launch (2.1.178, Jun 15). It closes a gap where a subagent could request a blocked action without review.
- Claude Code no longer treats scheduled tasks and webhooks as typing (2.1.183, Jun 19). In auto mode, they can no longer approve a pending action.
Worth your time
- Steering Claude Code (Jun 18): Anthropic lays out seven ways to instruct Claude Code. They include CLAUDE.md files, rules, skills, subagents and hooks.
- Opus 4.8 Build Day winners (Jun 17): 310 builders took part in a 12-hour hackathon on Jun 13. The winner, Tekton, rebuilds Tang Dynasty buildings in 3D.
- Built with Opus 4.7 winners (Jun 15): projects included medical training and electronics repair. First place went to Medkit.
That's the week. The next issue lands Sunday, Jun 28.
Reply and tell me what you're building. I read every reply.
— Sam