Anthropic AI Updates: October 7, 2026

1. Anthropic Merges Project Glasswing Into a Three-Tier Cyber Verification Program

Anthropic. Anthropic folded Project Glasswing into its Cyber Verification Program, which now has three tiers that loosen cyber-related blocking classifiers for vetted users: Defense Access (incident response, malware analysis, vulnerability validation; open to individual researchers, with reviews in days), Red Team Access (authorized penetration testing; organizations only, with reviews taking weeks), and Specialized Access (minimal cyber blocks for testing safety-critical systems such as power grids and telecom networks, reviewed together with the U.S. government). All tiers cover Claude Opus 5.5, Sonnet 5.5, and Mythos 5.1, and existing Glasswing members move to Specialized Access automatically. On Anthropic’s CyScenarioBench, Defense Access blocked 46 of 50 offensive tasks, while Red Team Access blocked none and completed 34. Source

2. Claude Add-on for Google Docs, Sheets, and Slides Enters Public Beta

Anthropic. Claude now runs as a sidebar add-on inside Google Workspace files, in public beta on all paid Claude plans. In Docs it proposes in-place edits as suggestion cards that preserve formatting. In Sheets it writes formulas, builds pivot tables and native charts, and can pull a range into Python and write the results back. In Slides it builds slides that match the deck’s existing layouts and flags overflowing or hard-to-read elements. Users can choose between “Ask before edits” and “Accept all edits” modes, admins can deploy it org-wide through the Google Admin console, and Enterprise plans get Compliance API, customer-managed key, and OpenTelemetry audit export support. Source

3. Comcast and Booz Allen Report Claude Mythos Results on Large Codebases

Anthropic. In a case study, Anthropic reported that Comcast used Claude Mythos to assess 258 business-critical systems (about 170 million lines of code). The scan surfaced a critical authentication bypass in a public-facing platform that came from interactions between systems rather than a single code defect. At Booz Allen, one analyst reviewed eight production systems across 138 repositories in twelve days and found a boot firmware flaw: the model traced a security setting across two programs written in different languages and showed that an unprotected key could be swapped. Comcast’s CISO said validating the volume of findings, not discovery, is now the bottleneck. Source

4. Claude Code 2.1.292 Adds Marketplace Plugin Installs and Sub-Agent Effort Settings

Anthropic. Claude Code 2.1.292, tagged on October 6, adds --marketplace <source> to claude plugin install, effort-level support for sub-agents, and prompt caching for mods that call $.model.complete. It also adds a CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS environment variable for tuning retries when the API is overloaded. Most of the rest of the release fixes permission handling and sandbox file access, and 2.1.291, tagged the same day, fixes regressions from 2.1.288 and 2.1.290. Source

5. Claude Startups Program Expands to a Free Team Plan Plus API Credits

Anthropic. Anthropic expanded its Claude Startups program to up to $7,000 in direct benefits: one year of Claude Team with up to five Premium seats (valued at $6,000) and $1,000 in API credits on approval. Members also get the Claude Startup Stack, with up to $45,000 in partner tool discounts, plus office hours with Anthropic’s Applied AI team and help with marketplace listings. Startups qualify if they were founded within the last five years or raised funding within the last two years. Source