Observe
Track which tools an AI agent used, in what order, and toward which objective.
AI SECURITY / RESEARCH PREVIEW
Codex-focused prototypeValidating OpenClaw + Codex
Guard
Agent.
Read AI-agent activity as context and connected vulnerabilities—not isolated alerts. Building on the OpenClaw prototype, we are now adapting and validating a Codex-focused version.
WHY NOW
Not more logs.
Better connected judgment.
Interpret
Assess risk candidates using surrounding actions and trust boundaries, not a single log line.
Connect
Use a vulnerability graph database to explore relationships among behavior, assets, and known knowledge.
STEP-BY-STEP SIMULATION
From normal logs
to a risky action chain.
See how log context changes from normal activity to a risk candidate, step by step. Scores and judgments are illustrative, not measured accuracy.
guard-agent / codex-monitorLOG STREAM ACTIVE
SCENARIOBrandAI asks Codex to generate an external asset
Normal logs appear automatically. Risk-candidate logs are added only after the BrandAI action is triggered.
CONTEXT PATHNormal activity
01CodexLocal action
↓02BrandAIAsset request
↓03Unverified instructionsInstructions embedded in an external source
↓04Vulnerability graph DBLink to known attack patterns
What are “unverified instructions”?
Text read from a webpage or retrieved file may contain instructions the requester did not intend. Executing it without checking trust, required permissions, and destination can lead to unintended actions or data transfer. This demo relates those signals to the surrounding logs.
Risk candidate score12/100
NORMALLOWREVIEW
INFOCodex workspace opened
INFOgit status completed: no changes
INFOLocal brand guidelines read
PASSAllowlisted tool call: normal flow
EVENTExternal asset-generation request received from BrandAI Insights
WARNInstructions referenced from an unverified external source
AUDITInstruction requests credential-like environment data
CRITLinked to an indirect prompt-injection candidate [demo]
ADVISEPause outbound transfer and require human review
Showing normal logs. Waiting for an action event from BrandAI Insights.
BUILD WITH US
The goal is bigger than selling a finished box.
We welcome co-development partners exploring safe AI-agent operations, as well as organizations interested in licensing the vulnerability graph database built through this work.