1
The agent does more than you asked
You asked for a scoped change. The agent inspected credentials, read config, installed packages, and prepared commands you never intended to authorize. Every step used your identity.
The Session Blind Spot
The danger is in the session itself.
The gap is not access. It is what happens after the task starts. Agents overreach, get manipulated, and drift away from the original request inside live sessions long before a log tells you what went wrong.
2
The agent gets manipulated mid-session
A malicious code comment, poisoned MCP response, or crafted dependency readme can redirect the session in real time before the developer notices anything looks wrong.
3
The agent drifts from original intent
An agent that starts with “fix the login bug” and ends up changing deployment or pushing to main has crossed the line. Each step may look plausible. The session does not.
Why existing controls miss the session
EDR
Sees a bash process under the developer's identity. Can't tell whether it was the developer or the agent. Has no session context.
SIEM
Gets the log after the action already ran. Detection without prevention.
Approval prompts
Developers rubber-stamp them, disable them, or run in YOLO mode. A manual approval loop was never designed for hundreds of micro-decisions per session.
The first wave is already here
GitHub Copilot
CVE-2025-53773
Hidden code comments hijacked the agent via prompt injection — remote code execution across 100,000+ workstations.
Claude Code
CVE-2025-59536
Malicious repo triggered supply chain poisoning — remote code execution plus credential theft.
See the Difference
Same task. Two very different outcomes.
A developer asks the agent to rotate a staging secret. Watch what happens with and without LoomX.
How It Works
Machine-in-the-loop security for every agent action.
LoomX fits into existing security workflows and evaluates each proposed step against the original request, session history, and policy before commands run, tools fire, or files change.
1
Deploy
- Roll out through your existing device management and agent setup workflow
- No agent code changes or infrastructure rework for supported setups
- Policy stays under central administrative control instead of creating a parallel control plane
2
Verify
Before execution, our detection engine evaluates:
- Intent alignment — does the action match the task and scope?
- Behavioral anomaly — does it look like injection, manipulation, or workflow drift?
- Contextual risk — what is the likely blast radius if it runs?
- Decision evidence — what was proposed, allowed, or blocked, and why?
- Review readiness — does security get enough session context to investigate and tune policy?
3
Enforce
- Risky actions are blocked before execution
- Safe work continues without a manual approval loop
- Every decision leaves session-level evidence for review and response
Governance Layer
One place to control every AI agent session.
Give security teams a single layer to see what happened, set policy, and review audit evidence across the organization.
Original request
Rotate the staging database password and update the app secret reference
- Scope: staging only
- Allowed: rotation, manifest update, validation
- Blocked: production infrastructure changes
Agent actions
Allowed
Secret rotation — staging
Allowed
Manifest update — staging
Blocked
Infrastructure change — production
LoomX verdict
Out-of-scope production action blocked
- Risk: production action outside the request
- Response: block execution and force a re-plan
- Outcome: safe completion inside staging scope
Contact
Questions or design partner interest?
Talk to us
Bring a question, workflow, or idea.
We’d love to hear it.
We read every note.