Agent execution security · Python systems · Operational infrastructure
I build and research systems that make AI-agent actions controllable, auditable and inspectable. My focus is the boundary between a model proposing an action and an operational system being authorized to execute it.
Core principle: reasoning authority is not execution authority. Approval, execution authority and evidence of an external effect are separate properties.
I lead Luftfiber AS, working with telecommunications and operational infrastructure in Norway. Production operations inform my approach to software: least privilege, explicit trust boundaries, durable state, failure recovery and verification of what actually changed.
I contribute technical reviews, reproducible tests and bounded conformance cases to public discussions on agent security, MCP/tool authorization and evidence sufficiency, including CoSAI / OASIS WS4 and agent governance interoperability.
My work emphasizes what evidence can actually establish. A missing observation is not automatically proof of a violation; a passing test is not proof of complete runtime enforcement. Participation in these discussions does not imply standards endorsement, certification or adoption of my projects.
| Project | What it addresses |
|---|---|
| REMORA Research | Policy-gated agent execution, exact-call and semantic authority, execution evidence, bounded evaluations and published negative results. |
| Agent Authority Conformance Profiles (AACP) | Vendor-neutral, draft conformance profiles for seven distinct authority and execution properties. Bounded Claim Reproduction (BCR), explicit inconclusive outcomes and no aggregate security score. |
| Assured Agent Execution | Reference control plane for exact-payload approval, separated execution roles, auditable lifecycle and selected system-of-record effect checks. |
| vericlaim | Evidence-linked CI checks that detect drift between documented claims, benchmarks and supporting artifacts. |
| pilotfish | Experimental policy-governed network link selection; operational router integration remains future work. |
- Research claims: machine-readable claim status, evidence level, sample sizes and caveats.
- Capability maturity: what is implemented, wired, persisted or still unestablished.
- Execution-path quickstart: the governed reference path and its credential-topology limits.
- Negative results: failed hypotheses, safety/utility trade-offs and open research boundaries.
- Federation and self-service evidence: pinned interoperability artifacts, reproduction levels and claim ceilings.
- Research paper: current manuscript and its evaluation limitations.
REMORA remains research/shadow-mode software. Bounded external reproductions of specific interoperability contracts do not establish production enforcement, complete runtime coverage, independent safety certification or universal agent safety.
I define a bounded property, identify the evidence needed to establish it, and build tests that can falsify it. Claims stay tied to specific revisions and assumptions. When evidence falls short, I narrow the claim rather than stretching the conclusion.
My engineering work spans Python, distributed systems, identity and access controls, policy-as-code, MCP/tool integration and network automation.
I welcome concrete engineering collaboration, independent reproduction and adversarial review of agent execution controls and evidence semantics. A useful starting point is a specific workflow: which action is proposed, who can authorize it, which credential path can execute it and how its effect can be verified.
Contact: support@luftfiber.no