Make the whole system visible.
Map models, retrieval, memory, agents, tools, identities, data stores, infrastructure, and operators as one security boundary.
Proposed / CONTROL EVIDENCETrust Boundaries for Tool-Using AI SystemsOpen research record +HASP LABS / AI SECURITY SYSTEMS
The model is not the system. Secure all of it.
HASP Labs helps teams building and deploying AI map system boundaries, engineer enforceable controls, test failure paths, and hand off evidence and runbooks operators can own.
- 01
Define what the system is allowed to trust.
Map every component, trust boundary, data path, identity, dependency, and recovery point before assigning authority or connecting tools.
- 02
Make security decisions enforceable.
Place identity, authorization, isolation, validation, approval, and evidence controls between model output and real effects.
- 03
Control how intelligence becomes action.
Make plans inspectable, delegate only the capability each task needs, preserve approvals, and keep cancellation and recovery available.
- 04
Test what the controls actually guarantee.
Exercise boundaries, enforcement, delegation, and recovery under adversarial and degraded conditions, with methods, limitations, and evidence visible.
RELATED RESEARCH / CONTROL EVIDENCE
Open the research behind each control.
Keep context attributable.
Track where retrieval and memory came from, what may use them, how they can change, and when they expire. Context can inform a decision; it cannot grant authority.
Published / CONTROL EVIDENCEWhere Retrieval Changes the Security BoundaryOpen research record +Keep identity outside the prompt.
Bind user, workload, agent, and tool identity independently of model context so every request has an attributable principal.
In progress / CONTROL EVIDENCEIdentity Continuity Across AI System BoundariesOpen research record +Put policy between decisions and effects.
Validate the operation, resource, arguments, budget, and approval at the enforcement point—and fail closed when evidence is missing.
Published / CONTROL EVIDENCETesting Policy Enforcement Between Models and ToolsOpen research record +Separate planning from permission.
Make proposed steps, dependencies, postconditions, and failure paths inspectable. A plan describes work; it does not authorize execution.
Proposed / CONTROL EVIDENCEHuman Approval as a Runtime Security BoundaryOpen research record +Narrow authority at every handoff.
Give each agent or task a named purpose, the minimum capability required, a bounded lifetime, and a revocable path back to its owner.
Published / CONTROL EVIDENCELeast-Privilege Delegation Across Multi-Agent WorkflowsOpen research record +Test controls at the point of action.
Exercise allowed, denied, malformed, unapproved, and degraded requests against the tool boundary. Verify that failures remain closed.
Published / CONTROL EVIDENCEFail-Closed Tool Gateways for Agent RuntimesOpen research record +Make effects attributable and recoverable.
Connect intent, identity, policy, delegation, approval, tool results, and recovery checkpoints in one causal evidence record.
Published / CONTROL EVIDENCEEvidence Contracts for AI Control DecisionsOpen research record +ONE SYSTEM / FOUR SECURITY DISCIPLINES
Trust is a system property.
Architecture defines the boundary. Engineering enforces it. Orchestration constrains action. Research tests what remains uncertain.SECURITY FOUNDATIONS
AI security depends on the systems beneath it.
Identity, infrastructure, hardening, cloud controls, supply-chain integrity, observability, compliance, and recovery determine whether AI controls can hold in production.- 01Identity & access
- 02Infrastructure hardening
- 03Isolation & least privilege
- 04Cloud & supply chain
- 05Observability & compliance
- 06Recovery & continuity
HOW WE WORK
From unclear risk to operable controls.
We map the system, build controls into the action path, exercise failure, and transfer the evidence and runbooks your team needs to operate it.See the operating boundary.
Make components, identities, authority, data movement, and recovery paths explicit.
Insert enforceable controls.
Put policy and evidence in the path where an AI system reads, decides, and acts.
Test controls and recovery.
Exercise decisions, actions, evidence, failure handling, and recovery before operators take ownership.
Leave an operable system.
Deliver controls, decisions, evidence, policy, and runbooks that the team can own.
AI SECURITY RESEARCH / OPEN LIBRARY
Test assumptions before they become dependencies.
Explore research on trust boundaries, retrieval, identity, policy enforcement, delegation, tool use, evidence, and recovery—with methods, limitations, maturity, and related work visible.Open the research index +HASP LABS / AI SECURITY BRIEF
Bring us the AI system you need to trust.
Discuss a system 3 steps / about 2 minutes / opens an email draftcontact@hasplabs.com