Offensive Security Digest

• Core flaw: Unauthenticated file upload in Joomla's com_baforms accepts any filename with no extension allow-list or CSRF, landing .php files in...

Two converging vectors are widening the offensive perimeter against AI systems.

• Supply-chain prompt injection now hides malicious instructions...

AI Agent Exploitation Vectors

• 🔥 Claude Code RCE: Proof-of-concept shows Claude Code auto-mode can be exploited for remote code execution during...

• UAF in rtmutex/futex PI since Linux 2.6.39 lets attackers reclaim stack memory and hijack kernel execution
• Achieves root privs and container...

• MDASH deploys 100+ AI agents across Windows code to hunt flaws at scale, surfacing 16 new issues including critical RCEs.
• Result: higher Patch...

AI now drives both concrete attacks and systemic risk expansion that offensive teams must monitor.

• Autonomous ransomware executed end-to-end by an...

Red teams face immediate exposure when using AI coding agents for reviews.

• RCE exploit hits Claude Code auto-mode during third-party library scans,...

Red teams must test LLM integrations for excessive agency—where models abuse backend APIs to perform unauthorized actions.

Key attack flow from the...

Zero-Day Closure

• 🔥 RoguePlanet Patched: Microsoft released an update to the Malware Protection Engine (version 1.1.26060.3008) fixing...

• June disclosure: Nightmare Eclipse published PoC for CVE-2026-50656, a race condition in Defender granting SYSTEM access on fully patched Windows...

ZDI reports a 490% spike in vulnerability submissions this month versus last April, overwhelming triage teams and shuttering programs like Internet...

Mythos-class models erode the assumed edge of Chinese hardware backdoors and manufacturer knowledge by accelerating vulnerability discovery for...

A classic Unix symlink attack named GhostApproval tricks AI assistants including Claude Code, Cursor, and Amazon Q into following deceptive links to...

0-Day Disclosures

• 🔥 ZDI-26-403 Ollama DoS: ZDI disclosed an unauthenticated denial-of-service in Ollama's downloadBlob function due to improper...

• CISA adds first AI agent platform (Langflow) to KEV with July 10 patching deadline after real credential theft campaigns
• DeepSeek AI independently...

AI Offensive Architectures

• 🔥 Cybersecurity AI Scientist: Chinese Academy of Sciences paper proposes the Cybersecurity AI Scientist framework...

Public zero-day dumps like Exploitarium force offensive teams to rapidly triage credibility before any testing or operationalization begins.

-...

AI agents are accelerating both offensive attacks and defensive research at machine speed.

• JadePuffer showed an agent fully executing ransomware:...

Cavern Manticore's .NET-based framework shows a clean modular design worth emulating for targeted operations.

• Red team emulation: Load modules into...

Kernel Exploit Development

• 🔥 Bad Epoll Linux Kernel Flaw: CVE-2026-46242 is a newly disclosed privilege-escalation vulnerability in the Linux...