01
WeeklyPROMPT INJECTION
Injection moved downstream
Retrieval, orchestration, and hidden instructions are the new weak point — not the model itself.
Read the injection seriesAI SECURITY INTELLIGENCE
Find the AI attack paths your team is about to ship
AIPwn tracks prompt injection, tool abuse, exposed endpoints, leaked secrets, and physical AI risks — before they turn into incidents.
Weekly notes, live exposure signals, and release-control thinking for teams building agents, copilots, RAG systems, and embodied AI.
THIS WEEK
This week's AI security signals
Real attack paths, reachable surfaces, and release risks — summarized for builders.
02
Deep diveTOOL ABUSE
Runtime permission is the blast radius
Fetch, shell, subprocess, and connectors decide how far a single prompt can reach.
See the tool-use research03
Live boardPUBLIC EXPOSURE
Exposure turns theory into incident
OpenClaw shows which endpoints, docs, and agent interfaces are reachable right now.
Open the watchboardTHREAT SURFACE
Where AI systems become pwnable
Five surfaces where AI risk becomes reachable.
01Input path
Prompt Injection
Hidden instructions, poisoned retrieval, and indirect control paths.
02Runtime
Tool Abuse
Shell, browser, file, connector, and API actions that expand blast radius.
03Secrets
Secret Exposure
Keys, logs, configs, tokens, and credentials leaked through AI workflows.
04Exposure
Public Surface
Open docs, unauthenticated endpoints, exposed agent interfaces.
05Physical AI
Physical AI
Robot, sensor, and action loops where model behavior reaches the real world.
METHOD
From signal to release decision
Every note follows the same loop — from a vague risk to a call you can act on.
Name the path
What is the actual exploit chain?
Expose the surface
Where is it reachable today?
Gate the release
Should this AI change ship?