Reported by 1 outlet
Claude for Chrome — Unpatched Flaw Lets Rogue Extensions Read Gmail, Calendar, and Docs
Why this score
- Breadth
- 0.50 / 2.0 · 1 source
- Authority
- 1.80 / 2.0 · best outlet weight
- Severity
- 4.90 / 5.0 · CVSS + content signals
- Velocity
- 0.00 / 2.0 · updates in 48h
- Pickup speed
- 0.00 / 0.6 · time to second source
- Recency
- × 0.00 temporal multiplier
- Community
- + 0.00 / 1.5 · reddit signal
- Score
- 0.0 / 10 = base × recency + community
Timeline
-
original reporting
Claude for Chrome — Unpatched Flaw Lets Rogue Extensions Read Gmail, Calendar, and Docs
securityweek.com · 91d ago · July 15th, 2026
Manifold Security has disclosed that Claude for Chrome v1.0.80 remains vulnerable to two flaws that allow any other browser extension with DOM access to claude.ai to trigger tasks that read the user's Gmail, latest Google Docs (including comments), and Calendar. The vulnerability persists eight versions after being reported to Anthropic on May 21, 2026.
The primary flaw: Claude for Chrome's content script listens for clicks on
#claude-onboarding-button, reads thedata-task-idattribute, and opens the side panel with the matching task. The handler never checksevent.isTrusted, so any extension can dispatch a synthetic click. Nine fixed task IDs remain in the allowlist — three of which (usecase-gmail,usecase-gdocs,usecase-calendar) read user data. In the default "ask before acting" mode, an approval dialog still fires (CVSS 7.7). In "Act without asking" mode, the task executes silently (CVSS 9.6).A secondary flaw: the side panel reads
skipPermissionsfrom its URL and boots intoskip_all_permission_checkswhen set. While not remotely exploitable today, a future bug exposing this parameter could enable fully silent data reads.Anthropic closed the reports as duplicates of the earlier ClaudeBleed issue but the code remains unchanged in v1.0.80 (byte-for-byte identical to v1.0.72). No CVE or advisory has been published.
Action: Disable "Act without asking" mode in Claude for Chrome. Audit browser extensions with claude.ai DOM access permissions. This is the third iteration of trust boundary issues in this extension.