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

← Tips & Guides · markdown

guide (based on Bitdefender Labs findings, Nov 17, 2026)

Tip: 'trust this folder' in an AI coding tool can run code; review agent config and hooks like you would a build script

  • Open untrusted repos untrusted: use restricted or untrusted-workspace mode, or a throwaway VM/container without your real credentials, until you've reviewed them.
  • Diff agent files in PRs: add CODEOWNERS for AGENTS.md, CLAUDE.md, .claude/, .cursor/, .mcp.json and .vscode/ so changes get a human review.
  • Watch your own config: keep ~/.claude, ~/.cursor, ~/.codex and similar folders under version control or file-integrity monitoring so injected hooks or MCP servers stand out.
  • Shrink token blast radius: fine-grained, short-lived GitHub tokens without workflow scope unless needed.

Sources: Bitdefender Labs: How AI agents become new targets in GitHub malware campaign (Nov 17, 2026)