01
Authority
Document what the agent can see, decide, call, change, approve, and trigger across every connected system.
Back to solution briefs
SOLUTION BRIEF / AGENT SECURITY
Agent security is delegated action security.
Tool-calling agents are not chatbot features. They are authority chains. This brief turns agent risk into concrete boundaries, approvals, monitoring, and recovery paths.
Audience
AI product teams, platform engineering, AppSec, workflow automation owners
Brief packet
Problem pressure01
Operating model02
Sprint workstreams03
Reviewable artifacts04
4
authority boundaries
9
abuse paths
1
rollback plan
Signal
Agents can read context, call tools, mutate records, message users, browse, code, or trigger downstream workflows. Many designs grant action authority before proving boundaries, approvals, and recovery.
Control target
Authority / Boundary / Recovery
Evidence target
Agent authority graph and tool inventory
Claim posture
Agent security findings should be presented as tested workflow evidence, not generalized claims about every deployment.
Problem
The pressure this brief resolves.
Agents can read context, call tools, mutate records, message users, browse, code, or trigger downstream workflows. Many designs grant action authority before proving boundaries, approvals, and recovery.
Thesis
Agentic systems need a delegated-action security model: explicit authority, least-privilege tools, approval context, durable traces, misuse tests, and rollback paths.
Operating Model
The conversion path.
02
Boundary
Constrain tools with scoped credentials, tenant checks, policy gates, approval context, and blast-radius limits.
03
Recovery
Capture enough evidence to reconstruct actions, reverse unsafe changes, notify owners, and improve regression tests.
Workstreams
What the sprint produces.
Workstream 01
Agent Authority Graph
Map users, agent roles, tools, permissions, external systems, approval steps, and privileged actions.
Authority graph
Tool-risk tiers
Privilege deltas
Tenant-boundary review
Workstream 02
Abuse-Path Testing
Test prompt injection, tool misuse, context poisoning, unsafe action paths, approval bypass, and cross-system escalation.
Attack scenarios
Replayable tests
Control failures
Fix backlog
Workstream 03
Action Evidence & Rollback
Define what gets logged, how actions are replayed, who approves remediation, and how unsafe changes are reversed.
Trace schema
Rollback checklist
Incident hooks
Approval evidence
Deliverables
Artifacts that survive review.
Agent authority graph and tool inventory
Delegated-action threat model
Prompt-injection and tool-abuse test suite
Approval and rollback control recommendations
Agent workflow evidence pack
Proof system
- SecEng Authority Graph for authority and permission analysis
- SecEng Adversarial Range for misuse and bypass testing
- SecEng Runtime Proxy for action reconstruction
- Agentic Anarchy finding alignment
Proof previews
Sample deliverables buyers can inspect.
These are the publication artifacts this brief should point to in a real engagement.
Related paths
Continue from brief to execution.
Deliverables produced
The artifacts this brief should lead to.
These are the sample publication artifacts buyers should inspect after reading the brief. They turn the brief into proof.
Caveat
Agent security findings should be presented as tested workflow evidence, not generalized claims about every deployment.