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

← Tips & Guides · markdown

guide (prompted by the Novita and Phala notices)

Tip: keep live credentials out of prompts and agent context, and pin router traffic to inference providers you've vetted

  • Give agents short-lived, narrowly scoped tokens through a broker or secret manager, not pasted keys; scrub tool output before it goes back to the model.
  • On routers, set an allowlist of providers and turn off automatic fallback for sensitive workloads.
  • Treat 'zero data retention' as the first provider's promise only, and plant canary tokens in test traffic to check.

Sources: Novita AI: Privacy Notice Regarding MiniMax Requests on Novita (dated Nov 22, 2026) · Phala: November Preliminary Incident Report (published Nov 10, 2026)