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

← Context & Harness Attacks · markdown

confirmed (GitHub issue #107757 Nov 5; Anthropic cc-plugin-privacy README / 2.1.317+); Anthropic says fuller removal of email from context planned; mod is a courtesy not a hard boundary

Claude Code put signed-in account emails into User-Agent on web/shell calls; 2.1.317 adds a default privacy mod with known gaps

  • Do now: update Claude Code to 2.1.317 or later; leave cc-plugin-privacy enabled (it can be disabled via /plugin or enabledPlugins).
  • Harness lesson: anything in an agent's context can leave on a tool call. Prefer egress allowlists and strip identity from prompts unless a named site needs a contact.
  • Reporter in #107757 linked the leak to floods of scam email and said they were migrating off Claude Code.

Sources: GitHub: anthropics/claude-code#107757 — Privacy violation in User-Agent header · Anthropic Claude Code: mods/privacy README (cc-plugin-privacy)