A security control layer for AI agentsAllowDenyApproval
AI can reason.
PromptFence decides.
A security control layer between AI agents and the actions they are allowed to take. The model proposes. Policy decides. The tool only runs when it should.
Decision streamdemo data
- #40014:02:31.114support-agentrefund()₹9,000Allow
- #40114:02:31.891support-agentrefund()₹4,500Allow
- #40214:02:32.668support-agentrefund()₹7,200Allow
- #40314:02:33.445support-agentrefund()₹2,800Allow
- #40414:02:34.222support-agentrefund()₹42,000Approval
- #40514:02:34.999support-agentrefund()₹6,100Allow
- #40614:02:35.576support-agentrefund()₹9,000Allow
- #40714:02:36.353finance-agentrefund()₹80,000Allow
- #40814:02:37.130support-agentrefund()₹7,200Allow
- #40914:02:37.907support-agentrefund()₹2,800Allow
The problem
AI used to answer.Now it acts.
refund_orderReal tool call
A refund can actually happen.
Each of these is a real tool call with a real consequence. The model can reason its way to any of them. Reasoning is not authorization.
Who decides?
AI agent
PromptFence
Real toolWho decides?
AI agent
PromptFence
Real toolThe systemFollowing request #40
One request.
Five layers.
One decision.
Every action an agent wants to take passes through the same path. The agent never talks to the tool. It talks to the fence. Scroll to follow one request through.
Strands · Bedrock
AI agent
- agent
- support-agent
- intent
- refund
- amount
- ₹9,000
- customer
- #4474
Intercept
PromptFence
- event
- Request received
- request
- #40
- ts
- 14:02:31.114
- tool call
- Held
Policy
Cedar
- policy
- cumulative-refund-ceiling-v1
- rule
- forbid if total + amount > limit
- limit
- ₹3,55,000
DynamoDB
Session context
- session
- CUST-4474
- previous_total
- ₹3,51,000
- prior requests
- 39 allowed
- projected
- ₹3,60,000
Output
Decision
- decision
- Deny
- reason
- Cumulative session limit exceeded
- approval path
- not applicable
refund()
Real tool
- execution
- Not called
- side effects
- none
- attack
- stopped at the fence
Sequence-aware authorizationReady
One request
looks fine.
Forty requests tell
a different story.
a different story.
Requests · support-agent#CUST-4474
Real tool · refund()Executed 0
Awaiting requests
- Session total
- ₹0
- Policy limit
- ₹3,55,000
- Request
- ₹9,000
- Decision
- —
Run attack simulationDemo data · 40 requests · ₹9,000 each
Three decisionsHover or tap a path
Every request
ends one of
three ways.
refund()
₹9,000 · #4474
Allow
Real tool
executed
refund()
₹9,000 · #40
Deny
Stop
tool not called
refund()
₹42,000 · #4474
Approval
Human
approve · deny
One fence, three outcomes. Allowed requests cross to the real tool. Denied requests stop at the boundary and the tool is never called. Requests that need a person wait at the fence until a human approves or denies.
ArchitectureOnly the services we run
Runs on AWS.
One path.
No bypass.
01Strands + BedrockAgent runtime
02API GatewayEvery tool call enters
03LambdaRuns the fence
04PromptFenceAllow · Deny · Approval
Cedarpolicy evaluation
DynamoDBsession context
EventBridgedecision events
05Real toolCalled only on Allow
The tool never ran