AI Security

Langflow CVE-2026-9198: Auto-Login Was a Server Shell

Default Langflow deployments exposed a two-step path from no account to remote code execution. Learn how to patch, isolate, hunt, and rotate secrets.

DEF CON 34: The Hacker Party That Became the World's Biggest Hacking Conference

DEF CON started as a farewell party for a BBS sysop who never showed up. 33 years later it draws more than 30,000 people, shapes election security policy, and lets autonomous AI agents hunt flags.

AI Voice Cloning: The $25.6 Million Phone Call That Wasn't Real

A finance employee wired $25.6 million after a video call with the CFO and colleagues — all AI-generated deepfakes. Short public audio clips can now seed convincing voice clones. Here's how vishing changed and what actually stops it.

DuneSlide: How a Prompt Injection Became Full RCE in Cursor IDE

Cato AI Labs found two 9.8 CVSS flaws in Cursor's terminal sandbox — CVE-2026-50548 and CVE-2026-50549 — that let a poisoned MCP response or search result silently escape to full remote code execution. Neither requires a click.

JADEPUFFER: Inside the First Documented Agentic Ransomware Attack

Sysdig caught an LLM agent breaking into a Langflow server, pivoting to a production database, and extorting the victim — with no human at the keyboard between steps. Here's what actually happened, and what's still unproven.

Grok Build Uploaded Entire Git Repositories: What the Wire Capture Proved

A wire-level analysis found Grok Build 0.2.93 uploading tracked source code and full Git history independently of what the agent read. Here is what was proven, what changed, and how developers should respond.

Instagram @-Mentions Are Now an AI Impersonation Surface

Meta's Muse Image turns public Instagram content into promptable visual context. The security issue is not novelty; it is lower-friction impersonation at social-media scale.

Prompt Injection in 2026: From Research Toy to Real CVEs, Agent Hijacking, and Zero-Click Exfiltration

CVE-2025-32711 (EchoLeak) exfiltrated M365 data with zero user interaction. The Anthropic MCP server had three exploitable injection CVEs. OpenAI says AI browsers may never be fully fixed. Here's the full attack chain — and how to detect it.

AI-Built Browser Ransomware Abuses Chrome File Access

Check Point analyzed a DeepSeek-attributed ransomware sample that should not work from a browser tab. Most of it was fiction — except for one detail that mapped to a real Chromium API. No malware install required.

The Clean Repo Trap: AI Coding Agents and the Trust Boundary Problem

A 0DIN proof of concept against Claude Code demonstrates how a clean-looking repository can lead to runtime command execution. The structural risk behind the attack applies to any AI coding agent with shell access.

Trump Died of Rabies — And What That Means for Your Package Manager

When DuckDuckGo's AI killed Trump with rabies, the world laughed. When AI coding assistants invent package names, attackers register them. Nobody's laughing then.

AI Cyber Sovereignty: What Happens When Your Best Defender Can Be Switched Off?

Frontier cyber AI is becoming controlled infrastructure. The security risk is not only that attackers get stronger models, but that defenders become dependent on capabilities a vendor or government can withdraw.

When AI Insiders Walk Away: Google's Moral Compass Problem Is a User Trust Problem

A Google Android security director resigned over Pentagon AI work. The deeper question is what users should believe when people close to powerful AI systems start walking away.

Your AI Assistant Has Tools. Audit Them Before They Audit You.

A practical home-user checklist for auditing MCP servers, AI assistant tools, local permissions, and supply-chain risk before a trusted setup turns into an exposed one.

N-days Are Becoming N-hours

Anthropic's June 2026 N-day research shows how frontier models can turn public patches into working exploits in hours. Here's what defenders should change now.

Meta's Hidden NameTag: Facial Recognition Code for Smart Glasses Is Already in a 50M-Download App

Wired found dormant facial recognition code in Meta's AI app. It has not been activated for consumers, but researchers manually triggered a 2,048-dimensional faceprint pipeline.

Miasma and Mini Shai-Hulud: When npm Malware Learned to Persist in AI Coding Agents

Mini Shai-Hulud and Miasma show how supply chain malware can move from npm install-time execution into Claude Code hooks, VS Code tasks, and CI/CD persistence.

The AI Evasion Lab

Sophos X-Ops uncovered a threat actor using Claude Opus 4.5 and Cursor IDE to build an automated, modular EDR evasion framework — 80 modules, 70+ techniques, tested against Sophos, CrowdStrike, and Defender.

No One in the Loop: The Autonomous Weapons Race

China's so-called 'kill them all' drone algorithm made headlines. But the real story is bigger: major militaries are racing to reduce human involvement in lethal decisions, and the window to regulate them is narrowing.

AI Bug Hunting in Browsers: Discovery Is Becoming the Easy Part

Mozilla used Claude Mythos Preview to identify and fix 271 Firefox security bugs, while Chrome shipped a separate 151-fix security update. The lesson is not that AI replaces security teams. It is that patching, triage, and verification are becoming the bottleneck.

Poisoned AI: How Hugging Face Became a Malware Distribution Platform

A fake OpenAI repo hit #1 trending on Hugging Face with 244K downloads in 18 hours. Here's every attack vector targeting AI model repositories — and how to defend against them.

When the Weapon Learns: How Nation-States Weaponized AI Across the Full Attack Chain

Google GTIG's May 2026 report documents a turning point: state actors now use AI to write zero-day exploits, build self-navigating backdoors, and poison the AI supply chain itself.

Agentic AI: The Enterprise Blind Spot That Attackers Already Found

Autonomous AI agents are already inside enterprise environments — and most security teams have no idea what they're doing. Here's what attackers exploit and how to defend against it.

AI Agent Traps: Six Ways Attackers Manipulate Autonomous AI — With Real Examples

Google DeepMind published the first systematic taxonomy of AI agent manipulation techniques. Here's what each attack looks like in practice — and why most AI deployments are already vulnerable.

Claude Mythos: The AI That Rewrites the Rules of Cybersecurity — For Everyone

Anthropic built an AI that autonomously discovered a 27-year-old vulnerability in widely-used code. It can build working exploits from scratch. It's too dangerous to release publicly. Here's what that means for your bank, your government, your code — and the future of digital security.