Archive — 2026-06Week 3

TechCrunch analysis on Trump administration's crackdown on Anthropic — "Who benefits?" reveals the takedown was triggered by Amazon finding a Fable 5 jailbreak. Published 2026-06-21, the piece revisits the U.S. government order — citing "national security concerns" — forcing Anthropic to take its two most capable models, Claude Fable 5 and Claude Mythos 5, offline, and names the trigger: Amazon researchers reportedly found a jailbreak vulnerability in Anthropic's Fable 5 model, and after Amazon CEO Andy Jassy raised it with the White House, the takedown order followed. The article argues the move is more politically motivated than substantively security-driven, and weighs who benefits: competitors may gain as Anthropic's restrictions level the field, yet Anthropic itself may benefit from looking powerful and willing to defy the government (downloads of Claude reportedly surged after prior clashes). Security experts had signed an open letter opposing the ban, arguing it strips "cybersecurity capabilities from network defenders," while critics say Anthropic contradicts itself by warning AI is too dangerous while shipping powerful models. It extends Bloomberg's 6/19 questioning of export-control law's reach, adding the "Amazon-reported jailbreak" origin and an industry-competition lens.

InfoQ: ~95% of Anthropic's internal analytics queries are now handled by Claude; after introducing analytics 'skills,' accuracy jumped from 21% to over 95% — InfoQ reported on 2026-06-21 that Anthropic now routes about 95% of its internal analytics requests through Claude, letting employees self-serve operational data without relying on the data team. The key was encoding analytics 'skills' for Claude (packaging business context, semantic definitions, and workflows): before such skills, Claude answered only 21% of questions correctly; afterward overall accuracy rose to over 95%, approaching 99% in specific domains. The article stresses that AI analytics is only as reliable as the underlying data platform, attributing the results more to data governance, semantic definitions, and operational discipline than to raw model gains; the shift freed the data science team to focus on more strategic work like causal modeling, forecasting, and machine learning.

Claude Code v2.1.185 released — stream-stall hint reworded: when the API response is delayed it now shows "Waiting for API response · will retry in …" (previously "No response from API · Retrying in …"), and the hint now triggers after 20 seconds of silence instead of 10.

Claude Opus 4.8 elevated error rates (resolved) — status.claude.com logged an incident affecting claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation of elevated Opus 4.8 error rates began 2026-06-20 17:07 UTC, a fix was implemented and monitored at 17:49 UTC, and it was marked resolved at 18:02 UTC — about 55 minutes. This continues the pattern of recurring frontier-model error-rate incidents since Opus 4.8 launched.

Bloomberg: Lutnick's Anthropic crackdown claims an unprecedented new power to regulate AI-model "usage," raising industry-wide legal questions — Bloomberg reported on 2026-06-19 that the Trump administration's pressure on Anthropic rests on an unprecedented use of export-control law (invoking the 2018 Export Control Reform Act): Commerce Secretary Howard Lutnick stretched the law's boundaries from governing the "transfer" of sensitive technology to controlling the very "usage" of frontier AI models — restricting who may access a model's API rather than the export of physical goods or software. Legal experts question whether the order holds up: a former senior Commerce official said it "might not restrict API/chatbot access at all," and warned export controls are being used as a "roving license to ban unsafe products." The piece notes the case could become a template for future government limits on any advanced AI model's user base, raising major legal questions for the whole industry about regulatory authority and the scope of export-control statutes.

Claude service disruption (resolved) — status.claude.com logged an "Elevated error rates on the Claude API" incident affecting claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation began 2026-06-19 08:17 UTC, a fix was implemented and monitored at 08:37 UTC, and it was marked resolved at 08:45 UTC — about 28 minutes. This was a general API error-rate incident (the status page named no specific model), the second of the day after the 06:07 Opus 4.8 incident, continuing the recent run of recurring Claude service error rates.

Nobel Chemistry laureate John Jumper leaves Google DeepMind for Anthropic — AlphaFold lead joins the Claude maker, aligning with its life-sciences push — Multiple tech outlets reported on 2026-06-19/20 that Google DeepMind VP and 2024 Nobel Chemistry laureate John Jumper announced he is leaving DeepMind after nearly nine years to join Anthropic. Jumper announced the move via a post on X on Friday 2026-06-19, saying he would take time to recharge before starting. He shared the 2024 Nobel Prize in Chemistry with DeepMind CEO Demis Hassabis for AlphaFold, the AI system that predicts the 3D structures of over 200 million proteins from genetic sequences. Neither Anthropic nor Jumper disclosed his role, but the hire aligns with Anthropic's expanding push into life sciences and computational biology. Bloomberg noted Jumper was also instrumental in Google's coding-tool efforts, an area Google has struggled to commercialize. The move follows Google engineering VP and Gemini co-lead Noam Shazeer's departure to IPO-bound OpenAI, underscoring talent drain from DeepMind.

Claude service disruption (resolved) — status.claude.com logged an incident citing elevated errors for the Claude Opus 4.8 model, affecting claude.ai, Claude Console (platform.claude.com), the Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation began 2026-06-19 06:07 UTC, a fix was implemented and monitored at 07:03 UTC, and the incident was declared resolved at 07:17 UTC, lasting about 70 minutes. It continues a recent string of incidents citing elevated error rates for Opus 4.8.

Claude Code v2.1.183 released — Improved auto mode safety: destructive git commands (`git reset --hard`, `git checkout -- .`, `git clean -fd`, `git stash drop`) are now blocked when you didn't ask to discard local work; `git commit --amend` is blocked when the commit wasn't made by the agent this session; `terraform destroy`/`pulumi destroy`/`cdk destroy` are blocked unless you asked for the specific stack. Added: a warning shown on stderr in `-p` print mode when the requested model is deprecated or auto-updated to a newer one (now also covering models set in agent frontmatter); an `attribution.sessionUrl` setting to omit the claude.ai session link from commits/PRs in web and Remote Control sessions; `/config --help` listing all shorthand keys for `/config key=value`. Changed: `/config` toggle now lets both Enter and Space change the selected setting, and Esc saves and closes instead of reverting; removed the startup "setup issues" line under the logo (use `/doctor` or `--debug`). Fixed: `thinking.disabled.display` "Extra inputs are not permitted" 400 errors on subagent spawns and session-title generation; WebSearch returning empty results in subagents; the terminal cursor stranded above the prompt after navigating history in vim mode with the native cursor; fullscreen TUI corruption (statusline mid-screen, duplicated spinner rows, merged text) in Windows Terminal under heavy nested-subagent load; turns silently completing with no output when the model returned only a thinking block (now re-prompts once); user-level skills appearing multiple times in slash-command autocomplete when multiple plugins are enabled; MCP servers requiring authentication exposing auth-stub tools to the model in headless/SDK mode; tmux teammate panes failing to launch under slow shell rc init and keystrokes during spawn leaking into the new pane; background tasks started by a teammate being killed when the teammate finishes a turn; scheduled task and webhook trigger deliveries being treated as keyboard input (now classified as task notifications, no longer able to approve a pending action or set the session title in auto mode); and focus mode showing "Ran N PostToolUse hooks" timing lines under each response.

Claude Code introduces Artifacts (public beta) — turn a session into a shareable, in-place-updating web page — On 2026-06-18 Anthropic announced Claude Code Artifacts in public beta on the claude.com blog: it turns the work from a Claude Code session into a live, shareable web page, drawing on full session context (codebase, connectors, conversation history). When Claude Code updates an artifact, the open page refreshes automatically at the same URL; version history allows restoration, and a gallery lets you browse all created artifacts. Use cases: PR walkthroughs (with diff and reasoning), filterable and sortable dashboards, release checklists that auto-populate as work progresses, incident pages combining timeline/commits/error metrics, license audits and security findings linked to code lines, and data-flow maps for privacy review. Private by default, shareable only with authenticated organization members; admins control access via org-level toggles and role-based scoping, with retention policies and compliance API visibility. Available in beta to Claude Team and Enterprise orgs, creatable from the Claude Code CLI and desktop app, with pages viewable in any browser.

Claude introduces enterprise-managed authorization for MCP connectors (beta) — Okta supported first, zero-touch connection on first login — On 2026-06-18 Anthropic announced enterprise-managed authorization in public beta on the claude.com blog: admins can provision MCP connectors for an entire organization once through an identity provider, and users get zero-touch access automatically on first login, with authorization configured centrally by the organization. It is the first implementation of the Enterprise-Managed Authorization extension to the Model Context Protocol, spanning Claude chat, Claude Code, and Cowork. Okta is the first supported identity provider (more coming soon); connectors supported at launch include Asana, Atlassian, Canva, Figma, Granola, Linear, and Supabase (Slack coming soon). Available in beta to Claude Team and Enterprise plan customers. Early adopters include HubSpot, Ramp, and Webflow — Ramp said that before enterprise-managed auth, onboarding a new hire meant a queue of per-connector OAuth approvals; now 2,000 employees are provisioned through Okta and fully connected on day one with zero extra steps.

Fortune: Anthropic CEO Dario Amodei has just one direct report; all other senior leaders report to his sister — Fortune reported on 2026-06-18 that Anthropic founder and CEO Dario Amodei directly manages only his chief of staff, while every other senior leader reports to president and cofounder Daniela Amodei, his sister. Amodei calls the arrangement "incredibly freeing," letting him focus on high-stakes operational matters (such as tense negotiations with the Pentagon over AI safety) rather than day-to-day management. The piece notes the unconventional structure reflects Anthropic's lower organizational complexity versus larger firms (e.g., Nvidia, with ~42,000 employees and a CEO with ~60 direct reports); Anthropic had roughly 2,300 employees at the end of 2025, is valued at $965 billion, and is preparing for an IPO.

JPMorgan follows Goldman Sachs in blocking Hong Kong staff from accessing Anthropic Claude — The Financial Times/Reuters reported on 2026-06-18 that JPMorgan Chase removed Claude models from the internal drop-down list of approved large language models (LLMs) for its Hong Kong employees, citing the wording of Anthropic's usage terms in its licensing agreement, which prohibit accessing Claude in China; Goldman Sachs had taken the same step for its Hong Kong staff back in April 2026. Although Hong Kong is not an officially supported region for Anthropic, international organizations had historically used Claude there via global service agreements thanks to the region's greater internet freedom relative to mainland China. The move comes amid rising US-China tensions over AI technology, data security, and controls on advanced computing, and as Anthropic negotiates with the White House to overturn its 'national security risk' designation (which drove the 6/12 export controls suspending Fable 5 and Mythos 5) while preparing for an initial public offering (IPO) — the Hong Kong access restrictions could further complicate its path to listing.

Anthropic publishes Project Fetch Phase Two — Claude Opus 4.7 autonomously controls a robotic quadruped, far faster than human teams — On 2026-06-18 Anthropic published Project Fetch Phase Two, redoing its 2025 experiment to assess whether newer Claude models could control a robotic quadruped autonomously without human help. Results: on the four tasks completed by all participants, Claude Opus 4.7 averaged 37.7 times faster than the non-Claude human team and 18.9 times faster than the Claude-assisted team, and about 20 times faster than the original experiment's fastest human team overall; it also generated roughly one-tenth the code of the previous Team Claude (1,045 lines vs 10,309). The model excelled at sensor connectivity, programming-interface development, and object detection (finishing in about 9–12 minutes), but still struggled with precise ball manipulation — the core 'fetch' objective requiring nuanced real-time control. The researchers see this as early-stage 'physical agentic AI,' where models shift from assisting humans to operating tools autonomously, paralleling the trajectory of software AI coding capabilities.

Claude services disruption (resolved) — status.claude.com logged an incident affecting claude.ai: impact began at 06:55 UTC on 2026-06-18, the cause was identified and a fix was underway at 07:41 UTC, impact ended at 07:40 UTC, and it was resolved at 08:19 UTC, lasting about 45 minutes. This was a general service disruption (the status page named no specific model), distinct from the recent string of incidents that specifically cited Opus 4.8 error rates.

Korea JoongAng Daily: Anthropic executive says Fable 5 / Mythos 5 access could return 'in coming days' — Korea JoongAng Daily reported on 2026-06-18 that Anthropic's Managing Director of International, Chris Ciauri, said at a press conference at Conrad Seoul on 2026-06-17 (the Seoul office opening) that 'we are very confident that in the coming days, the models will become available again.' It was the company's first optimistic restoration timeline after the US government's 2026-06-12 national-security export-control directive forced Anthropic to disable Fable 5 and Mythos 5 for all customers to comply. Anthropic has been in near-daily talks with the administration since last week; its other models (such as Opus 4.8) are unaffected.

Claude Code v2.1.181 released — Added: `/config key=value` syntax to set any setting from the prompt (e.g. `/config thinking=false`; works in interactive, `-p`, and Remote Control), `sandbox.allowAppleEvents` opt-in to let sandboxed commands send Apple Events on macOS, and `CLAUDE_CLIENT_PRESENCE_FILE` env var to suppress mobile push notifications while you're at the machine. Changed: bundled Bun runtime upgraded to 1.4, long paragraphs now stream line-by-line (no longer waiting for the first line break), API drops mid-thinking now auto-retry (no more 'Connection closed while thinking'), the subagent panel auto-hides idle subagents after 30s (caps at 5 rows with scroll hints), the MCP OAuth browser page matches Claude Code's style and auto-closes on success, and fullscreen URL opening now requires Cmd/Ctrl+click. Fixed: prompt caching not reading on custom `ANTHROPIC_BASE_URL` and Foundry due to a per-request attestation token, Write/Edit producing 0-byte or truncated files on network/cloud-synced drives, `open`/`osascript`/browser auth failing with error -600 on macOS (added Apple Events entitlement), startup crash on corrupted null project entries in `.claude.json`, macOS TUI freezing at session start during Spotlight reindex (Ctrl+C unresponsive), foreground subagents spawning unbounded nested chains (now capped at 5 levels), and `/recap` and forks using the previous model after a model switch.

Claude Developer Platform launches Workload Identity Federation (GA) — short-lived credentials replace static API keys — On 2026-06-17 Anthropic announced that Workload Identity Federation (WIF) for the Claude Developer Platform is generally available (GA) on the claude.com blog: it replaces static Anthropic API keys with short-lived, scoped credentials. The system verifies the workload's signed OIDC token, matches its claims against your federation rules, and issues a short-lived access token bounded by service account roles, so developers never have to handle a static API key when building (avoiding the risks of key creation, rotation, and leakage). It is compatible with any OIDC-compliant identity provider, including AWS IAM roles, GCP and Kubernetes service accounts, Azure managed identities, GitHub Actions tokens, and Okta. It covers all Claude API endpoints, including when accessed through the first-party SDKs and Claude Code.

Elevated error rate for Claude Opus 4.8 (resolved) — status.claude.com logged an incident affecting claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code, and Claude Cowork: investigation of elevated Opus 4.8 errors began at 15:34 UTC on 2026-06-17, a fix was deployed and monitored at 16:01 UTC, and it was resolved at 16:28 UTC, lasting about 54 minutes. It was the fourth such incident on 6/17 after those at 00:47, 05:08, and 08:24, continuing the pattern of recurring frontier-model error spikes since Opus 4.8 launched.

Fortune/AP: A New York House primary becomes an AI-industry political battle, with Anthropic-aligned groups spending over $10 million to back Bores — On 2026-06-17 Fortune, citing the Associated Press, reported that New York State Assemblyman Alex Bores is running in the June 23 Democratic primary for a Manhattan-based U.S. House seat (NY-12). Bores is the lead sponsor of the RAISE Act, which would require major AI firms to file safety reports on 'catastrophic' risks affecting more than 50 people (e.g., AI failures at nuclear plants or engineered viruses). A super PAC backed by OpenAI investors, 'Leading the Future' (funders include OpenAI president Greg Brockman, Marc Andreessen, and Joe Lonsdale), spent over $7.6 million on ads against Bores; groups aligned with Anthropic and its investors, such as 'Public First' (led by Brad Carson), put in over $10 million in support, while crypto billionaire Chris Larsen pledged an additional $3.5 million. The race highlights Silicon Valley's split over AI regulation — Anthropic leans toward stricter state-level rules, while the OpenAI camp opposes state restrictions and prefers a unified federal standard.

Anthropic opens Seoul office and announces partnerships across Korea's AI ecosystem — In a 2026-06-17 announcement, Anthropic named KiYoung Choi as Representative Director and signed an MOU with Korea's Ministry of Science and ICT, working with the Korea AI Safety Institute on model safety evaluation and cybersecurity research and to promote safe public-sector AI adoption. Enterprise rollouts: NAVER adopted Claude Code across engineering teams, Nexon uses it for live-service game development, LG CNS is deploying Claude to thousands of employees, Hanwha Solutions is adopting it via AWS Bedrock with data-residency compliance, and Samsung SDS is deploying Claude across Samsung Electronics for knowledge work and software development. Startup/product integration: Channel Corp is embedding Claude into Channel Talk, its customer-AI platform serving 230,000+ companies across Korea, Japan, and the U.S. Academic/nonprofit: a partnership with the National AI Research Lab (NAIRL) gives 60 researchers Claude access, and Good Neighbors Korea uses Claude to ease social workers' administrative load. Developer community: Claude for Startups is now live in Korea, a Claude Build Day was held with BASS Ventures (100+ attendees), and a Push to Prod hackathon is planned with Replit and Korea Investment Partners. This is the actual opening following the 5/26 appointment of a Korea representative, advancing Anthropic's APAC footprint.

Claude Opus 4.8 elevated error rate (resolved) — status.claude.com logged an incident affecting claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code and Claude Cowork: investigation of elevated Opus 4.8 error rates began at 08:24 UTC on 2026-06-17, the issue was identified and a fix deployed at 08:54 UTC, monitoring began at 09:37 UTC, and the incident was resolved at 10:03 UTC, lasting about 1 hour 39 minutes. It is the third incident the same day after the 00:47 and 05:08 UTC incidents on 6/17, continuing the pattern of recurring frontier-model error incidents since Opus 4.8's launch.

Major Claude Design update — design-system imports, brand-guideline auto-correction, two-way handoff with Claude Code, and a fix for its token-burning problem — Multiple tech outlets reported on 2026-06-17 that Anthropic Labs significantly updated Claude Design: users can import one or more design systems from a GitHub repo, design files, or uploads, and Claude builds with those components, checks its output against the system, and auto-corrects before showing results; admins can lock a single design system for company-wide consistency. A new WYSIWYG canvas adds drag/resize/align editing, and a single click hands off a design to a Claude Code project, carrying design tokens and component structure to scaffold a full app; the terminal supports /design and /design-sync to create, edit, import/export, and sync design systems from a local codebase. Exports now include PDF and PowerPoint, with integrations for Adobe, Canva, Gamma, Lovable, Miro, Replit, Vercel, and Wix. Claude Design is now in beta on all paid plans (Pro, Team, Enterprise); usage now shares one limit pool with chat, Claude Cowork, and Claude Code (instead of a smaller separate pool), easing earlier complaints about rapid token consumption. Access via claude.ai/design or the Claude desktop sidebar.

Claude Opus 4.8 elevated error rate (resolved) — status.claude.com logged an incident affecting claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code and Claude Cowork: investigation of elevated Opus 4.8 error rates began at 05:08 UTC on 2026-06-17, a fix was deployed and monitoring began at 05:44 UTC, and the incident was resolved at 06:37 UTC; the error window ran 04:59–05:41 UTC (9:59pm–10:41pm PT on 6/16), about 42 minutes. It continues the pattern of recurring frontier-model error incidents since Opus 4.8's launch, adding another the same day after the 00:47 UTC incident on 6/17.

Claude Sonnet 4.6 and Opus 4.8 elevated error rate (resolved) — status.claude.com logged an incident affecting claude.ai, Claude Console (platform.claude.com), Claude API (api.anthropic.com), Claude Code and Claude Cowork: investigation of elevated Sonnet 4.6 and Opus 4.8 error rates began at 00:47 UTC on 2026-06-17; by 01:52 UTC Sonnet 4.6 success rates had recovered while work continued on Opus 4.8; the incident was resolved at 02:50 UTC, lasting about 2 hours 3 minutes. It continues the pattern of recurring frontier-model error incidents since Opus 4.8's launch, adding to the multiple incidents on 6/16.

Claude Opus 4.8 elevated error rate (resolved) — status.claude.com logged an incident affecting claude.ai, Claude API (api.anthropic.com), Claude Code and Claude Cowork: Opus 4.8 request error rates rose from 20:45 UTC on 2026-06-16, investigation began at 20:50 UTC, a fix was deployed and monitored at 21:01 UTC, and it was resolved at 22:02 UTC (impact window ~20:45–20:58 UTC, ~13 minutes). This was the third Opus 4.8-related error incident of the day, continuing the pattern of recurring frontier-model error spikes since Opus 4.8's launch.

Claude Code v2.1.179 released — Fixes: mid-stream connection drops now preserve partial responses (no raw error, spinner no longer stuck at 'running tool'); WSL2 mouse-wheel scrolling regression (Windows Terminal/VS Code, introduced in 2.1.172); Linux sandbox `denyRead`/`allowRead` glob performance over large directory trees (which had bloated the Bash tool description and made sessions unusable); single-digit replies after a turn mistakenly captured as session ratings; multiple promo banners stacking on the welcome screen (now one promo per session); Ctrl+O not showing a subagent's transcript when viewing it; clicking the prompt input not returning focus from the subagent/footer panel; remote-session background tasks stuck as 'still running' between turns. Improvement: plugin loading performance in remote sessions.

Claude Opus 4.8 brief elevated error rate (resolved) — status.claude.com logged an incident affecting claude.ai and Claude Code: on 2026-06-16, Opus 4.8 error rates were elevated from 19:41 UTC (12:41 PDT) to roughly 19:53 UTC (12:53 PDT), about 12 minutes; a monitoring notice was posted at 19:59 UTC and the incident was resolved at 20:16 UTC. This was the second Opus 4.8-related error incident of the day.

Elevated error rate across multiple Claude models (resolved) — status.claude.com logged an incident affecting claude.ai, Claude API (api.anthropic.com), Claude Code and Claude Cowork: investigation of elevated errors across multiple models began at 17:29 UTC on 2026-06-16 (covering Sonnet, all Opus versions and Haiku 4.5, peaking around a 10% error rate); an initial fix was deployed and monitored at 18:00 UTC, the issue was identified as concentrated on Opus 4.8 and Haiku 4.5 at 18:15 UTC, another fix was deployed and monitored at 19:22 UTC, and it was resolved at 19:32 UTC, lasting about 2 hours 3 minutes in two phases (first all Sonnet/Opus, then Opus 4.8). This was the day's first and largest incident, followed by two more brief Opus 4.8 error events the same day.

Fortune: US government shutdown of Anthropic models ignites a global scramble for sovereign AI — Fortune reported on 2026-06-16 that after the US government fully cut off worldwide access to Fable 5 and Mythos 5 (described by officials as the first use of an AI "kill switch"), politicians across many countries voiced concern over dependence on US tech infrastructure, fueling calls for sovereign AI (national control of models, compute infrastructure, and data). UK MP Al Carns noted British hospitals, companies, and researchers all lost Fable 5 access; former security minister Tom Tugendhat said national security has shifted from "cannons" to "code"; AI minister Kanishka Narayan called AI capability the core national-security question. French former PM Édouard Philippe called AI critical infrastructure that "others can unplug," and 2027 presidential candidate Bruno Retailleau called it a "wake-up call." In the European Parliament, Finnish MEP Aura Salla warned Europe cannot rely on tech stacks "switched off overnight by a foreign government," and German MEP Sergey Lagodinsky urged pooling compute with middle powers via a consortium; European Commission President Ursula von der Leyen said the EU "cannot afford to depend on others for the technologies that keep our hospitals running." The EU's European Technological Sovereignty Package (unveiled 6/3) targets €422 billion over a decade across semiconductors, data centers, cloud, and AI infrastructure, with the Cloud and AI Development Act aiming to triple EU data-center capacity within 5–7 years.

Claude Code v2.1.178 released — added `Tool(param:value)` permission rule syntax to match a tool's input parameters (with `*` wildcard), e.g. `Agent(model:opus)` to block Opus subagents; skills in nested `.claude/skills` directories now load (shown as `<dir>:<name>` on name clashes), and the nested `.claude/` agent/workflow/output-style closest to the working directory wins on collisions; improved auto mode (subagent spawns are evaluated by the classifier before launch); improved `/doctor` layout and Remote Control error messages; `/bug` now requires a description before submitting. Fixed several subagent issues (transcript viewing, message dropping, backgrounding), compaction not honoring `--fallback-model`, requests still failing with auth errors after credential refresh, background sessions stuck showing 'Working', `claude agents` workers failing with 401, and Esc dismissing the CJK IME candidate window canceling the task in VS Code.

Anthropic negotiates with US Commerce officials to restore Fable 5 / Mythos 5 access — Globe and Mail reported on 2026-06-15 that senior Anthropic technical staff met with Commerce Department officials in Washington that Monday, with National Cyber Director Sean Cairncross attending the working-level meeting; the two sides have met almost daily since the government first reached out the previous Friday (6/12). The government wants assurances the models 'cannot be used to harm the U.S.,' fearing deployment by military-intelligence users in 'countries of concern' such as China and Russia; Anthropic argues the alleged jailbreak found only 'minor' flaws that other public models can also find, and that it tested Fable 5 with the government and received deployment approval before release. Commerce Secretary Howard Lutnick has held regular calls and is attending the G7 in Évian-les-Bains, France; over 80 cybersecurity executives (including leaders from Nvidia and Adobe) signed an open letter in support, urging the restrictions be lifted. No restoration timeline as of publication.

Anthropic hit with class-action lawsuit over allegedly overstated Claude Max usage limits — multiple tech outlets reported on 2026-06-15 that Karl Kahn, a Washington DC resident, filed a proposed class action against Anthropic in US federal court that day, alleging Anthropic overstated the usage allowances of its Claude Max plans: the actual caps for the Max 5x and Max 20x tiers are 'far below the advertised amount' and are hard to determine and opaque. The plaintiff says that after upgrading to the $200/month Max 20x plan in April 2025 he quickly hit weekly limits, with a single five-hour session consuming about 15% of his weekly allowance. The suit seeks class-action status on behalf of all US consumers who purchased Max subscriptions since the plans launched in April 2025; Anthropic declined to comment. It echoes the runaway Max-plan token consumption complaints reported by Decrypt on 6/10.

Claude Opus 4.8 elevated error rates incident (resolved) — Anthropic's status page logged an incident affecting claude.ai, Claude Console, Claude API, Claude Code, and Claude Cowork: investigation of elevated Opus 4.8 error rates began at 06:20 UTC on 2026-06-15, a fix was implemented and monitored at 08:14 UTC, and the incident was resolved at 08:56 UTC, lasting about 2 hours 36 minutes.