Anthropic AI Updates: April 25, 2026

1. Anthropic Details 2026 Election Safeguards Ahead of US Midterms

Anthropic. Anthropic published an update on how it is preparing Claude for the 2026 US midterms and other major elections worldwide, covering political-bias evaluations, usage policies, and election-information features. Pre-launch bias evaluations scored Opus 4.7 at 95% and Sonnet 4.6 at 96% for impartial engagement with prompts across the political spectrum, and the company has open-sourced the evaluation methodology and dataset for third-party replication. Anthropic also flagged ongoing work with The Future of Free Speech at Vanderbilt, the Foundation for American Innovation, and the Collective Intelligence Project on freedom-of-expression behaviors, alongside automated classifiers and a threat-intelligence team enforcing election-related Usage Policy violations. Source

2. Anthropic Postmortem Traces Claude Code Quality Complaints to Three Separate Harness Bugs

Anthropic. Anthropic published a postmortem attributing the past month of “Claude is dumber” reports in Claude Code, the Agent SDK, and Claude Cowork to three independent changes — none of which touched the API or inference layer. On March 4, the team lowered Claude Code’s default reasoning effort from high to medium to fix UI-freeze latency, then reverted on April 7 after user pushback; on March 26, a bug caused idle-session “thinking pruning” to fire every turn instead of once, making Claude seem forgetful (fixed April 10); and on April 16, a verbosity-reducing system prompt instruction degraded coding quality on Sonnet 4.6, Opus 4.6, and Opus 4.7, reverted April 20 in v2.1.116. All three are now resolved, and Anthropic is resetting usage limits for all subscribers as of April 23. Source

3. Project Glasswing Uses Claude Mythos Preview to Hunt Cross-OS Vulnerabilities

Anthropic. Anthropic launched Project Glasswing, an AI-vulnerability-discovery initiative built on an unreleased Claude Mythos Preview model that has reportedly turned up thousands of high-severity bugs, including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg bug missed by 5M automated test runs. Anthropic is committing up to $100M in usage credits and $4M in donations, restricting access to roughly 50 organizations including AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. The program positions Claude as a defensive security tool rather than an offensive one, with vetted partners agreeing to coordinated disclosure terms. Source

4. Claude Code v2.1.119 Lands Persistent /config, Custom PR URL Templates, and Hooks Timing

Anthropic. Claude Code v2.1.119 makes /config settings — theme, editor mode, verbose, and similar — persist to ~/.claude/settings.json and respect project, local, and policy override precedence, ending the long-standing complaint that interactively-set preferences vanished on restart. The release adds a prUrlTemplate setting that points the footer PR badge at custom code-review URLs (e.g., internal Gerrit or Phabricator) instead of github.com, broadens --from-pr to accept GitLab merge-request, Bitbucket pull-request, and GitHub Enterprise URLs, and lets --print mode honor agent frontmatter tools:/disallowedTools: to match interactive behavior. Hooks PostToolUse and PostToolUseFailure inputs now include a duration_ms field for tool execution timing, PowerShell tool commands gain Bash-style auto-approval in permission mode, and a new CLAUDE_CODE_HIDE_CWD env var hides the working directory in the startup logo. Source