PPhying中文
Curated security research, vulnerabilities, advisories and tools for practitioners.

From prompt to pwned: chaining LLM and web bugs to Admin

blog.quarkslab.com | research | #ai-security | #web-security | #red-team | #llm | #pentest | #insecure-output-handling
📄 Raw .md⬇️ Download .md🔗 Original ↗

Summary

A Quarkslab red-team case study: insecure output handling in an LLM medical assistant let them chain bugs from a low-privileged account to admin takeover — trusting the model was the first domino.
Published
Collected

Related coverage

research · vercel.com
DeepsecBench: evaluating model performance in finding cybersecurity vulnerabilities
Today we're releasing DeepsecBench, a benchmark that evaluates how well different models find security vulnerabilities in application code.
research · vercel.com
Security boundaries in agentic architectures
A framework for drawing security boundaries in agentic architectures. Most agents run with zero isolation between the agent and the code it generates. Learn where to draw the boundaries, from secret injection to full application sandboxing.
blog · vercel.com
The three types of AI bot traffic and how to handle them
Bots account for over 20% of all web traffic. A quarter of that is AI crawlers alone. But not all bots are bad. Some are key to how your site gets discovered. Learn how AI bots shape visibility and what happens if you block them.
research · github.com
php2shell: Pure HTTP Full-Chain RCE and ASLR Bypass on Stock PHP 8.4 Configurations
Overcomes historical constraints where PHP heap exploitation required display_errors=On, restricted memory limits, or target application vulnerabilities, demonstrating reliable pure-HTTP RCE and ASLR defeat against standard production ini configurations.
research · github.com
Infinity Core: Linux Kernel Core Subsystems LPE and Container Escape Research
Unlike most modern Linux kernel exploits that depend on networking, eBPF, or unprivileged user namespaces, these vulnerabilities reside in fundamental kernel core subsystems, rendering standard container sandboxes and Android application domain restrictions ineffective against local privilege escalation.
research · projectdiscovery.io
How abliterated models can get you pwned
# How abliterated models can get you pwned Oct 6, 2026 · Prince Chaddha Any open model whose weights have been edited can carry a backdoor, whether it's a task-specific fine-tune, a merged adapter...
Back