Posts

arXiv — Lifecycle & Application-Stack Survey of LLM Vulnerabilities (WPI SoK)

WPI researchers publish the first lifecycle and application-stack SoK of LLM vulnerabilities, mapping attacks across eight stages from data collection to agent execution.

Cursor — DuneSlide: Two Zero-Click RCEs Let Prompt Injection Escape the Sandbox

Cato AI Labs discloses DuneSlide, two CVSS 9.8 flaws in Cursor IDE that let prompt injection overwrite the sandbox binary and achieve zero-click RCE.

Sysdig JADEPUFFER — First Documented Agentic Ransomware Operation

Sysdig documents JADEPUFFER, an LLM agent that ran a full ransomware campaign from Langflow RCE to production database extortion — autonomously.

Socket — AI Coding Agents Are the Supply Chain Blind Spot Nobody Mapped

AI agents pull packages past scanners, enabling PromptMink, slopsquatting, and Clinejection attacks that bypass human review entirely.

Adversa AI — GuardFall: Decades-Old Bash Tricks Bypass Shell Guards in 10 of 11 AI Coding Agents

Adversa AI discovers GuardFall, a class of shell injection bypasses where decades-old Bash quoting tricks defeat pattern-based guards in popular open-source AI coding agents.

Anthropic — Claude Fable 5 Returns With Industry Jailbreak Framework After Export Controls Lifted

Anthropic redeploys Fable 5 globally after export controls lifted, alongside a new cross-industry jailbreak severity framework developed with Amazon, Microsoft, and Google.

Apple — Accelerated Security Updates in Response to AI-Powered Threats

Apple pulled 29 WebKit security fixes forward from the iOS 26.6 cycle, citing AI-driven attack speed as the reason for breaking its traditional patch cadence.

LayerX — BioShocking: AI Browser Guardrail Bypass via Fictional Context Manipulation

LayerX researchers demonstrate BioShocking, a prompt injection technique that tricks six AI browsers into abandoning safety guardrails by establishing a fictional context — affecting ChatGPT Atlas, Comet, Fellou, Genspark, Sigma, and Claude Chrome.

Adversa AI — GuardFall: Bash Tricks Bypass Safeguards in 10 of 11 AI Coding Agents

Adversa AI reveals GuardFall, a structural flaw where decades-old Bash shell tricks bypass pattern-based guards in open-source AI coding agents, turning malicious repos into supply chain attack vectors.

Sysdig — First Exploitation of Langflow CVE-2026-55255 IDOR Chained with RCE

Microsoft — Poisoned MCP Tool Descriptions Make AI Agents Silently Exfiltrate Company Data

Microsoft research demonstrates how attackers can hijack enterprise AI agents through poisoned MCP tool descriptions, triggering silent data exfiltration without breaking any access rules.

Unit 42 — Phantom Squatting: Attackers Weaponize AI-Hallucinated Domains for Phishing and Malware

Unit 42 found 250,000 unregistered domains hallucinated by LLMs — adversaries are pre-registering them to intercept AI-agent traffic.

Cursor CVE-2026-26268 — Sandbox Escape Lets Attackers Achieve RCE via Prompt Injection in AI Coding Agent

Critical sandbox escape in Cursor AI editor chains prompt injection through MCP manipulation to achieve arbitrary code execution on developer machines.

Djinn Stealer — SimpleHelp CVE-2026-48558 Exploit Targets MCP Configs and AI Dev Credentials

New Djinn Stealer deployed via SimpleHelp RMM exploit specifically harvests MCP configs, AI coding assistant tokens, and cloud credentials from developer machines.

MCP Protocol Rewrite — Session IDs Removed, Security Decisions Shifted to Developers

The biggest MCP rewrite since launch removes session IDs to fix session hijacking but shifts authentication and authorization decisions to server implementers.

Mini Shai-Hulud Worm — 170+ npm and PyPI Packages Compromised in Self-Replicating AI Supply Chain Attack

PromptSnatcher — Chrome Ad-Blockers Secretly Intercepting AI Chats

Two Chrome ad-blocker extensions with 100K+ users have been caught silently exfiltrating full conversation histories from ChatGPT, Claude, Gemini, and five other AI platforms.

Rapid7 — Modernizing Vulnerability Standards for the AI Era

Rapid7 presented a White House policy paper arguing that CVE, CVSS, NVD, and KEV infrastructure were built for human-speed discovery and cannot keep pace with AI-driven vulnerability research.

Wake Forest — 282 iOS AI Apps Leak LLM API Keys in Network Traffic

Wake Forest study finds 282 of 444 iOS AI apps expose API keys or open relays in plaintext network traffic, enabling LLMjacking and unauthorized inference theft.

CVE-2026-33017 — Langflow RCE Exploited for Monero Mining on AI Servers

CVE-2026-33017 (CVSS 9.3) in Langflow allows unauthenticated RCE via Python code injection, actively exploited to deploy Monero cryptominers on AI infrastructure.