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

← AI · markdown

vendor research (Malbear Labs, Nov 8, 2026, from an exposed attacker workspace); outcomes are Malbear's reading of the transcripts, not confirmed by victims

Operation Snowball: an exposed workspace shows a jailbroken DeepSeek Flash agent running a two-week intrusion into a European telecom with almost no human input

  • Why it matters: the hunt can't look for a human operator's working hours or session timing; look for the agent's fast, repetitive, parallel activity instead.
  • Patch: BroadWorks (Cisco's August fix for CVE-2026-20320; keep OCI port 2208 off the internet), EOL JBoss AS 6, Liferay 7.0, and the KEV entries in its toolkit (CVE-2022-22536 SAP ICM, CVE-2023-4966 CitrixBleed, CVE-2024-3400 PAN-OS).
  • Lessons from the transcripts: signature-only AV (here OfficeScan with behavior monitoring off) was routed around in one session; BladeLogic RSCD (port 4750) reachable from app servers gave SYSTEM.
  • See the hunt item for indicators.

Sources: Malbear Labs: Operation Snowball: Inside an Autonomous AI Intrusion Campaign (Nov 8, 2026) · Infosecurity Magazine: Jailbroken AI Agent Ran Two-Week Attack on European Telecom