GitLost: How a Single GitHub Issue Can Trick AI Agents Into Leaking Private Repos
Researchers at Noma Labs disclosed GitLost, a prompt-injection flaw that let a single crafted GitHub Issue trick AI-powered Agentic Workflows into leaking private repository contents publicly, using a...
DuneSlide: Critical Zero-Click RCE Bugs in Cursor IDE Put Fortune 500 Developer Machines at Risk
Two critical zero-click RCE vulnerabilities (CVE-2026-50548, CVE-2026-50549) in Cursor IDE, dubbed DuneSlide, allow attackers to escape the AI coding agent sandbox via prompt injection with no user interaction...
Microsoft Warns: Claude Code GitHub Action Exploitable via Prompt Injection to Leak CI/CD Secrets
Microsoft Threat Intelligence disclosed a prompt injection flaw in the Claude Code GitHub Action that allowed attackers to access /proc/self/environ and steal API keys from CI/CD runners. Anthropic...
OpenAI Launches ChatGPT Lockdown Mode to Block Prompt Injection Data Exfiltration
OpenAI has released ChatGPT Lockdown Mode, a new security feature that disables outbound network capabilities to cut off data exfiltration pathways exploited in prompt injection attacks. Available to...
Google Gemini Voice Assistant Hijacked via WhatsApp, Slack and SMS: Researchers Bypass All Google Defenses
SafeBreach researchers demonstrate how attackers can silently hijack Google Gemini through malicious payloads in WhatsApp, Slack, SMS, and other messaging app notifications, bypassing all of Google patched defenses...
AI Supply Chain Attack: 575+ Malicious Skills on Hugging Face and ClawHub Deliver Trojans, Cryptominers, and AMOS Stealer
Threat actors have uploaded 575+ malicious AI skills to ClawHub's OpenClaw ecosystem and abused Hugging Face repositories to deliver trojans, cryptominers, and AMOS Stealer. The campaign uses indirect...