[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Context & Harness Attacks · markdown

vendor research (Forever Security, Nov 4); researcher test only, no attacks in the wild reported; Anthropic acknowledged the report as valid (per the researchers) and says Dispatch is closed to new users

ThisBash: a browser extension that can script claude.ai could drive Claude Desktop through Dispatch and run code on the computer

  • Preconditions: an extension (malicious, compromised by an update, or buggy) that can run scripts on claude.ai, and a user signed in to claude.ai and Claude Desktop with Dispatch on and the desktop app open.
  • Why approvals didn't help: the approval prompts showed in the web page the extension controlled, so the script clicked Allow every second.
  • Do now: turn off cross-device Dispatch if you don't need it; keep personal Claude accounts with Dispatch off work machines; restrict which extensions may run on claude.ai; watch agents for unexpected package installs and new outbound connections.

Sources: Forever Security: ThisBash Escape, Claude Dispatch flaw enables browser extension RCE on the OS (Nov 4) · The Hacker News: malicious browser extensions could use Claude Dispatch to run code (Nov 6) · The Register: Anthropic won't fix a Claude Dispatch flaw (Nov 5)