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

← AI · markdown

confirmed research (SOCRadar Threat Research Unit, Nov 4; actor attribution high confidence for a single operator)

SOCRadar: 'KeyHunter' operator used Hermes Agent and OpenClaw bots to harvest 2,233 LLM API keys, map 247 open LLM gateways and run fake student sign-ups at 20 universities

  • Keep secrets out of anything a web server serves: check that /.env*, /wrangler.toml, /.vercel/output/config.json, /_next/env.js, /.streamlit/secrets.toml, /fly.toml and similar paths return 404 on your deployments, and look at Wayback Machine snapshots too, since the actor checked old copies.
  • LLM keys: set spend caps and alerts, scope keys per app, and rotate any key that has ever been in a repo or a client bundle.
  • Self-hosted gateways (one-api, new-api, LiteLLM and similar): require auth on every route, including model lists and admin panels.
  • IOCs: 209.112.91[.]74 (port 9997 dashboard), 173.231.39[.]244, 123.60.2[.]245, 14.47.118[.]194, 66.42.94[.]67; domains senhai[.]store, astrdark[.]cyou, xxtt[.]me, sajuai[.]xyz, vvbin[.]me. Full list in the report.

Sources: SOCRadar: KeyHunter Operation: Harvard Alumni Data, Stanford Applications and LLM Key Theft (Nov 4)