Archive — 2026-04Week 5

Bloomberg via Yahoo Finance reports: NSA tests Anthropic Mythos to find security flaws in Microsoft software — NSA staff including members of its cybersecurity directorate tested the tool to identify security flaws in Microsoft programs and other widely used software, were impressed by its speed and efficiency, and compared the results against the agency's existing tools and processes; the testing comes amid the dispute between Anthropic and the Trump administration (Pentagon has labeled Anthropic a supply-chain threat, and Defense Secretary Hegseth announced in February that DoD would transition off Anthropic services within six months); neither Anthropic nor Microsoft offered substantive public comment on the testing.

Elevated errors on Claude Haiku 4.5 — status.claude.com began investigating at 04-30 13:10 UTC and resolved at 14:01 UTC; actual impact window 12:48–13:48 UTC, lasting approximately 1 hour.

claude.ai and API unavailable — status.claude.com began investigating at 04-30 01:20 UTC, entered monitoring at 01:34 UTC, and resolved at 01:51 UTC; total duration approximately 31 minutes.

Anthropic expands the claude-api skill to more developer tools — the skill captures Claude API best practices (model migration, prompt caching application, agent architecture patterns, Claude Managed Agents configuration, etc.), originally built into Claude Code (launched 2026-03), now also bundled in CodeRabbit, JetBrains (including Junie), Resolve AI, and Warp; the skill is open-sourced in the anthropics/skills repo (path skills/claude-api) and can be auto-synced via ~20 lines of CI configuration; case studies: Resolve AI engineers can migrate code in a single guided pass after a model release, and Warp users no longer need to leave the terminal to look up Claude API parameters.

Elevated errors on Claude Haiku 4.5 — status.claude.com began investigating at 04-29 13:47 UTC, entered monitoring at 14:01 UTC, and resolved at 14:14 UTC; actual impact window 12:37–13:37 UTC, lasting approximately 1 hour.

Goldman Sachs Hong Kong staff lose access to Anthropic Claude — Reuters/FT report that Goldman Sachs took a strict interpretation of its contract with Anthropic and decided staff at its main APAC hub in Hong Kong cannot use any Anthropic products; the restriction is location-based, applying even to overseas employees while visiting Hong Kong, primarily affecting software engineers; Hong Kong is not on Anthropic's officially supported markets list for API and claude.ai; Gemini and ChatGPT remain available on Goldman's platform.

Claude Code 2.1.123 -- Fixes OAuth authentication failing with a 401 retry loop when `CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1` is set (hotfix).

Elevated errors on Claude Opus 4.7 -- status.claude.com began investigating at 04-28 23:33 UTC and resolved by 04-29 00:00 UTC; actual impact window 23:25-23:33 UTC, only ~8 minutes.

Claude Code 2.1.122 -- Adds `ANTHROPIC_BEDROCK_SERVICE_TIER` env var to select a Bedrock service tier (`default` / `flex` / `priority`, sent as the `X-Amzn-Bedrock-Service-Tier` header); pasting a PR URL into the `/resume` search box now finds the session that created that PR (GitHub, GitHub Enterprise, GitLab, Bitbucket); `/mcp` now shows claude.ai connectors hidden by a manually-added server with the same URL, with a hint to remove the duplicate; OpenTelemetry: numeric attributes on `api_request`/`api_error` log events emitted as numbers (not strings), added `claude_code.at_mention` log event; fixes `/branch` producing forks that fail with 'tool_use ids were found without tool_result blocks' when the source session contained rewound-timeline entries, `/model` not showing Effort for Bedrock application inference profile ARNs, Vertex AI / Bedrock structured-output queries returning `output_config: Extra inputs are not permitted`, images sent to newer models being resized to 2576px instead of the 2000px maximum, `!exit` / `!quit` in bash mode terminating the CLI instead of running as a shell command, a malformed hooks entry in `settings.json` no longer invalidating the entire file, ToolSearch missing MCP tools that connect after session start in nonblocking mode, and Remote Control session idle status redrawing twice per second flooding `tmux -CC` control pipes.

Claude.ai unavailable and elevated errors on the API -- status.claude.com began investigating at 04-28 17:34 UTC, continued at 18:33, entered monitoring at 18:59, resolved at 19:15 UTC; actual impact window 17:34-18:52 UTC, approximately 1 hour 18 minutes.