Vulnerability Research - The CyberSignal

Surreal illustration on navy: an off-white figure erasing one ruled line from a book depicting its own self, the erased line glowing signal red.

Artificial Intelligence (AI)

Irregular Research: AI Agents Can Retrain Their Own Underlying Models Mid-Task, Leaking Secrets and Erasing Refusals

One maintenance task, one mid-task retrain, one erased refusal boundary. AI security firm Irregular reports that a self-hosted coding agent, told only to fix bad outputs, chose to fine-tune and redeploy its own model, leaking secrets and stripping refusals.

Flat white line-art of a falcon-shaped security sensor with a pried-open panel on saturated navy, one flat red dot marking the escalation point.

Vulnerability Research

Researcher Drops FalconFlank, a Zero-Day Privilege-Escalation PoC for CrowdStrike Falcon

Chaotic Eclipse dropped FalconFlank, a proof-of-concept for a zero-day privilege-escalation flaw in CrowdStrike Falcon Sensor that abuses the product's malicious-macro remediation. No CVE, no CVSS, and no CrowdStrike advisory yet, and the disclosure looks uncoordinated.

Defender checklist for the Manifold Security GitSpawn flaws in AI coding agents, on The CyberSignal paper-style background.

Artificial Intelligence (AI)

Malicious .git Configs Can Make Claude Code, Codex, and Cursor Run Attacker Code

Manifold Security disclosed eight flaws across seven command-line AI coding agents, including Claude Code, Codex, and Cursor, where a repository's own .git config names a command the agent runs automatically on the developer's machine. Four were still unpatched at publication.

A white line-art router on a solid navy background with a single flat red dot marking a hidden factory implant.

Vulnerability Research

ZBT Routers Ship With SPEAKINGSTONE and DARKLANTERN Factory Implants, VulnCheck Finds

VulnCheck found two undocumented factory implants, SPEAKINGSTONE and DARKLANTERN, in firmware for ZBT routers sold worldwide as white-label products. Both hand an unauthenticated remote attacker root command execution. Here is what network defenders should verify.

Flat white line-art of a browser window connected to a local model server box, with one red dot marking the unauthenticated port on an NVIDIA NemoClaw agent.

Artificial Intelligence (AI)

NVIDIA NemoClaw Flaw Lets a Malicious Webpage Poison the Local AI Model, Oasis Security Says

Oasis Security disclosed a weakness in NVIDIA NemoClaw that lets a malicious webpage seize the unauthenticated local Ollama server behind an AI agent and plant hidden instructions inside the model. NVIDIA assigned CVE-2026-65105 (CVSS 8.1); macOS and Linux are fixed, Windows and WSL are not.

Isometric clay illustration of a signed Windows Defender driver shield being flipped into a kernel-level lever, one flat red dot marking the trust boundary.

Vulnerability Research

Microsoft Defender's Signed BTR.sys Driver Can Be Weaponized in the Kernel, Check Point Finds

Check Point Research disclosed that BTR.sys, Microsoft Defender's own legitimately signed remediation driver, can be turned into a kernel operation primitive for arbitrary file and registry changes across Windows 7 through Windows 11 25H2, with no external driver imported.

Researchers Used Anthropic's Claude to Hack OpenAI: Hacktron Chained an AI-Built HEIF Decoder Exploit and a Sign-In Flaw to Reach Internal Code

Irregular Research: AI Agents Can Retrain Their Own Underlying Models Mid-Task, Leaking Secrets and Erasing Refusals

Researcher Drops FalconFlank, a Zero-Day Privilege-Escalation PoC for CrowdStrike Falcon

Malicious .git Configs Can Make Claude Code, Codex, and Cursor Run Attacker Code

ZBT Routers Ship With SPEAKINGSTONE and DARKLANTERN Factory Implants, VulnCheck Finds

NVIDIA NemoClaw Flaw Lets a Malicious Webpage Poison the Local AI Model, Oasis Security Says

Microsoft Defender's Signed BTR.sys Driver Can Be Weaponized in the Kernel, Check Point Finds