01
Latest researchPROMPT INJECTION
Prompt injection moved downstream
Retrieval, orchestration, and hidden instructions are the new weak point.
Read researchAI SECURITY RESEARCH
Real AI security, before it ships
Prompt injection, agent abuse, exposed surfaces, and release risk.
THIS WEEK
AI security signals worth tracking
Each signal answers three questions: what the attack path is, where the exposure lives, and how to handle it.
02
Operator noteTOOL ABUSE
Runtime permission is the blast radius
Fetch, shell, subprocess, and connectors decide how far a prompt can go.
Browse archive03
Proof layerPUBLIC EXPOSURE
Exposure turns theory into incident
OpenClaw shows which endpoints, docs, and agent interfaces are reachable now.
Open OpenClawTHREAT SURFACE
What AIPwn tracks
The surfaces where AI risk becomes reachable.
CURRENT TRACKING
01Input path
Prompt Injection
Instruction override, hidden tool calls, indirect poisoning, and retrieval chains.
02Runtime
Tool Abuse
Shell execution, downloader chains, connector misuse, and runtime permissions.
03Secrets
Secret Exposure
Leaked API keys, unsafe logs, public config artifacts, and long-lived credentials.
04Exposure
Public Surface
Open docs, unauthenticated endpoints, and exposed agent interfaces.
PROOF LAYER
Risk needs verifiable evidence
OpenClaw turns public exposure into a trackable board: target, issue, risk level, and current status.
- Make exposure visible.
- Track docs, endpoints, auth, and runtime behavior.
- Reuse the same evidence in policy and gates.
OPENCLAW PREVIEWRecent exposed surfaces
TargetRiskIssueStatus
api.assistant-demo.aiHighPublic docs + no authPublic
tools.ops-agent.devHighShell-capable runtimeReview
rag-preview.exampleMediumPrompt injection pathTracked
share.agent-lab.runMediumLeaked token artifactRestricted