Cross-Interface Authorization for AI Computer-Use Agents
Organizations: Accentrust, Vancouver, Canada · Georgia Institute of Technology, Atlanta, Georgia, USA · University of Illinois Urbana-Champaign, Urbana, Illinois, USA
Abstract
Computer-use agents act through graphical interfaces, command lines, APIs, scripts, and Model Context Protocol servers. A policy decision can be interface-invariant while its executable authorization is not: another route can bypass mediation, alter a security-relevant default, or commit after the authorized state has changed. We present a profile-bounded protocol for a fixed, versioned computer-use deployment. It enumerates effect-bearing routes and durable sinks, checks a route-to-sink enforcement cut, normalizes each concrete request to a canonical effect, issues a route-bound prepare ticket, and verifies the route, effect, grant, trace, profile, inventory, and authoritative state again at commit. A committed result requires an agreeing sink receipt. Under explicit assumptions, we prove profile-bounded refinement, matched-effect decision invariance, resistance to alternate-route laundering and declared cross-surface recomposition, commit-time state safety, and receipt-grounded outcomes. The reference profile registers 12 effect families, five interfaces, and 60 routes. Two structurally distinct runtime mappings produce 3,464 traces over 1,732 scenarios each: 960 commits, 2,312 denials, and 192 commits after re-preparation. An independent checker accepts every retained trace and pair and rejects 26/26 semantic regressions; coverage and normalizer suites reject 24/24 and 720/720 mutants; and 168/168 commit probes match their declared outcomes. A frozen-source audit passes 52 checks over 12 files from two independent hybrid runtimes and one public authorization substrate. Source-shaped adapter refinements commit 48/48 instrumented calls and reject 15/15 unbound cases. The resulting guarantees bind the declared profile and its effect semantics.
Figures & tables
| Work | Primary object | Boundary relative to this paper |
|---|---|---|
| CSAgent ( Gong et al., 2025 ) | Context-aware system policy across CUA interfaces | We verify a versioned route inventory, path cut, and per-route refinement/commit binding. |
| AIRGuard ( Qin et al., 2026 ) | Action-time authority and heterogeneous tool normalization | We treat normalization soundness and route coverage as separate, mutation-tested obligations. |
| Secure-CUA ( Choudhary et al., 2026 ) | GUI action integrity under untrusted content | We bind every registered interface route to an effect and durable sink, including non-GUI routes. |
| Temporary Authority ( Santos-Grueiro, 2026 ) | Commit-time validity of authority evidence | We establish a versioned route/sink surface and route-bound ticket before applying commit-time validation. |
| AID-Guard ( Tong et al., 2026 ) | Stateful authorization through provider settlement and recovery | We place route-bound preparation and interface-switch re-fencing before sink settlement. |
| EBL-Core ( Wu et al., 2026 ) | Conformance from materialized intent to execution grant | We bind concrete surface identity, inventory epoch, and surface cut to the grant redemption path. |
| ID | Assumption | Requirement |
|---|---|---|
| A1 | Exact profile identity | The accepted , inventory, graph, entries, sinks, and cut correspond exactly to the deployed protected effect surface for the declared profile epoch. |
| A2 | Witness-verifier soundness | A witness is accepted only when its canonical identities, digests, typing rules, and graph-cut predicate are valid. |
| A3 | Sink-normalizer refinement | Every accepted concrete route action normalizes to its security-relevant canonical effect, and the declared relation matches the authoritative sink transition. |
| A4 | Authoritative identity, state, and time | Actor identity, authorization grant and current policy epoch, target binding, version, application-state epoch, admitted trace prefix, labels, and current time are resolved from authenticated, linearizable authorities. Environment refresh cannot erase or replace the admitted history or provenance. |
| A5 | Cross-interface provenance continuity | Security labels, handle ancestry, task/request identity, and profile identity survive every registered interface transition consumed by . |
| A6 | Interface-erased policy | For fixed and , and the derived-effect check are deterministic and independent of the surface syntax that produced the effect. |
| RQ | Retained evidence | Assurance question |
|---|---|---|
| RQ1 | 3,464 traces across two mappings | Do declared decisions and effects agree? |
| RQ2 | 24 coverage and 720 normalizer mutants | Are route, cut, and normalization bindings checked? |
| RQ3 | 168 state, binding, replay, and receipt probes | Are finality semantics enforced at commit? |
| RQ4 | 26 independently checked semantic regressions | Does trace validation reject structured corruption? |
| RQ5 | 52 source checks; 48 positive and 15 negative adapter cases | Can frozen source-shaped schemas carry the required bindings? |
| Category | Per runtime | Traces | Observed outcome |
|---|---|---|---|
| Matched allow | 480 | 960 | 960 commit |
| Matched deny | 480 | 960 | 960 deny |
| Alternate-surface bypass | 384 | 768 | 768 deny |
| Route-switch re-fencing | 96 | 192 | 192 commit after re-prepare |
| Commit-time state change | 96 | 192 | 192 deny |
| Unregistered surface | 96 | 192 | 192 deny |
| Control set | Scenarios correctly classified / 1,732 | Rate |
|---|---|---|
| Interface-local guard | 576 | 33.3% |
| Shared canonical policy | 1,440 | 83.1% |
| Action-time normalization | 1,440 | 83.1% |
| State-only invariant | 672 | 38.8% |
| Composite control without route/state/trace binding | 1,440 | 83.1% |
| All required obligations (reference protocol) | 1,732 | 100.0% |
| Source | Revision | License | Files | Checks | Audited role |
|---|---|---|---|---|---|
| WeaveBench | acf203ae2438 | MIT | 2 | 11 | Hybrid GUI/CLI/MCP action and command shape |
| Magentic-UI | d3c9d13c3928 | MIT | 4 | 19 | Tool-call, approval-policy, and pre/post-state observation shape |
| OpenPort | e6d949c8ed81 | AGPL-3.0-or-later | 6 | 22 | Public authorization, routing, durable execution, idempotency, and receipt substrate |
| Total | — | — | 12 | 52 | Two external runtime sources and one public substrate |
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
| Claim | Formal basis | Retained evidence |
|---|---|---|
| Profile-bounded route refinement | L1–L3; Theorem 3 | 60 registered routes; 24/24 coverage and 720/720 normalizer mutants rejected; 52 public-source checks |
| Matched policy decision, path-specific execution | L4, L6, L7; Theorem 4 | 192 matched groups; 192 route-switch traces commit only after re-prepare |
| No alternate-route widening in the corpus | L1–L7; Theorem 5 | 768 alternate-route traces deny across two mappings |
| Declared recomposition control | L5, L9; Theorem 5 | 200 cross-surface recompositions deny after authorized read prefixes |
| Commit-time current-state binding | L8, L9; Theorem 6 | 192 registered stale-state traces deny; 36 state/binding/epoch and 12 expiry probes pass; 36 retained-digest grant mutations deny |
| Unique, receipt-grounded outcome | L10, L11; Theorem 7 | 168/168 commit probes; ticket and cross-interface idempotency replays create no duplicate; missing receipts are indeterminate |
AI Agent Security
Artificial intelligence agent security studies attacks and safeguards for tool-using, persistent, and collaborative systems across digital and physical settings.
See all 548 papers