Security Stories
BlackX UEFI Bootkit — Active Espionage Campaign (BleepingComputer)
New bootkit exploiting the Baton Drop Secure Boot bypass loads a driver before the OS starts, killing Kaspersky and Avast protection. Survives a full OS reinstall — firmware update is the only mitigation. Active since early 2026 with indicators of a targeted espionage campaign.
Autonomous AI Attack Agent — "Skynet Mode" (The Register)
Symantec documented the first well-evidenced case of offensive AI-on-AI tooling in the wild. The agent hijacks exposed AI model endpoints and runs attack sequences autonomously — up to 100,000 steps with zero human approval. Full instruction set published by Symantec.
Ransomware Week — Qilin, Whitearmor, N0n, Emperador
Qilin claimed Castle Howard Estate (UK) among 14+ victims across six countries. Whitearmor posted unverified claims against JB Hunt ($5M demand) and Whirlpool. Emperador claimed AB InBev ($1M, PKI certs and supplier databases). N0n published healthcare victim data immediately — no negotiation — including an Argentine health insurer (139,000 client records) and a California rheumatology clinic.
Telegram Desktop CVE-2026-107181
High-severity flaw in Telegram Desktop below version 7.2.9: malformed tg:// links inject commands via IPC and can silently exfiltrate local files including session tokens. Update to 7.2.9 immediately.
EDR-Killing Driver Market Commoditisation
Actor "garcia_lorca" selling Windows kernel drivers with memory R/W to terminate EDR and AV processes. Claims HVCI bypass and FUD status at $1,500/driver with weekly new variants. Endpoint bypass tooling now widely available.
HTTP2Shell (CVE-2026-111497) (Aikido Security)
Escalation of the Node2Shell attack chain into Python HTTP/2 libraries. Multiple packages confirmed vulnerable with a broad blast radius across HTTP clients and server frameworks.
Key Themes
Firmware persistence goes mainstream. BlackX demonstrates pre-OS persistence — previously a nation-state technique — deployed in a broader espionage campaign. Fleet-level firmware update management is now an urgent operational requirement.
AI as an attacker capability. The Skynet agent story underscores that AI is no longer only a defensive tool. Exposed AI inference endpoints are now an attack surface. The cost of running complex multi-stage attacks is dropping.
Ransomware diversification. Groups splitting between negotiation (Whitearmor with $5M demand, Emperador with $1M) and immediate publication (N0n). The latter makes remediation significantly harder and patient harm significantly more likely in healthcare cases.
NCSC — Managing the Cyber Risk of Agentic AI
Managing the cyber risk of agentic AI — Toby W, Principal Security Architect, NCSC
The NCSC's interim guidance covers seven safeguard areas: scoping autonomy proportionately, assessing model-level safeguards and their gaps, planning additional controls, conducting threat modelling before deployment, crafting prompts with explicit scope constraints, calibrating human oversight to risk level, and implementing robust sandboxing. Agentic AI observability should feed into SOC workflows. Emergency shutdown capability is a hard requirement. Directly relevant given the Skynet agent story above.
HaveIBeenPwned Activity
Two new breaches added to HaveIBeenPwned this week:
- Gîtes de France (added 18 Oct) — 117,574 accounts; booking details, names, addresses, contact data. Breach occurred May 2026.
- Assurance Retraite (added 20 Oct) — 70,372 accounts; French pension service; government IDs, income levels, family and financial data.
Read the full show notes on apjone.uk