QenSAi · AIAGF · Governance in UseVisual guide · updated 16 October 2026
The visual guide
Everything we show, on one map.
Follow one request from the moment a person asks to the moment a person decides. Read the story, check the evidence, then open any stage to watch how it runs.
6 · Mapped to the OWASP Top 10 for Agentic Applications
Where each risk in the OWASP list (published 2025-12-09) is addressed in the story, and how strong the evidence is today.
ASI01 Agent Goal HijackThe job's rules and expected outcome are fixed in the oversight record before anything runs; the truth gate checks results against them.Recorded: an approved correction changed the next response (15 Sep 2026); a person redirected an agent that had made an unfinished experiment its goal, and the correction was kept in the oversight record (9 Oct 2026, record-keeping only).RecordedOpen cited evidenceStage 1Stage 3Stage 4Stage 7
ASI02 Tool Misuse & ExploitationEvery measure must point to a real governing rule and registered function; each handoff carries a Two-Value Pair check; agents act only inside the frame.Designed. Tool and MCP server hardening is not shown.Partial: designed or explainedOpen cited evidenceStage 4Stage 5Stage 6
ASI03 Identity & Privilege AbuseNo credential, no entry: the front gate checks who is asking and whether they may; a person authorizes before work runs.Recorded: a person authorized before anything ran (15 Sep 2026). Workload-identity standards such as SPIFFE or OIDC for agents are not shown.RecordedOpen cited evidenceStage 2
ASI04 Agentic Supply Chain VulnerabilitiesOnly registered functions and canonical records count; files carry a governed identity from creation.Designed. Third-party packages and registries are outside what we show.Partial: designed or explainedOpen cited evidenceStage 4Stage 9
ASI05 Unexpected Code Execution (RCE)Staged work is not authorized work: only accepted, authorized tasks reach a worker.Designed for authorization. Sandboxing and network isolation are outside our layer.Partial: designed or explainedOpen cited evidenceStage 2Stage 6
ASI06 Memory & Context PoisoningThe context an agent works from is the oversight record and canonical records, carried with the task, not recalled from the agent's own memory.Recorded in part: the task contract and its records now travel in the agent's first message (2 Oct 2026).Recorded, in partOpen cited evidenceStage 3Stage 4
ASI07 Insecure Inter-Agent CommunicationAgents talk through a governed relay; questions and answers stay tied to the same request on the Response Table.Recorded in part: one desk carried one request to three specialist agents, with a person arming each request (2 Oct 2026). Detecting unsanctioned side channels is not shown.Recorded, in partOpen cited evidenceStage 6Stage 8
ASI08 Cascading FailuresThe truth gate stops a chain when a claim of completion is not backed by evidence; governance state cannot jump from healthy to hostile without passing intermediate states.Recorded: the gate caught a false "done" and stopped the chain, then exposed its own gap when it counted an honest "no data" reply as completed (1 Oct 2026, two local runs); the gap was fixed on 2 Oct 2026.RecordedOpen cited evidenceStage 7
ASI09 Human-Agent Trust ExploitationA reply is not completion. The person decides on the Response Table from the record, and every page labels what kind of claim it makes.Recorded: completion is now checked against criteria; a claim with no evidence fails (2 Oct 2026).RecordedOpen cited evidenceStage 7Stage 8
ASI10 Rogue AgentsEach request is authorized by a person, results are checked against the record, and credentials lose validity as governance state degrades.Designed. Containment of an agent that has escaped its environment is outside our layer.Partial: designed or explainedOpen cited evidenceStage 2Stage 7
Outside our layer: Sandboxing and network egress control; Protection of binaries against reverse engineering; Containment of agents that have escaped their environment. Qensai governs authority and evidence; it works alongside containment, not in place of it.
Reading the labels
Every page says what kind of claim it makes. Read the limits each page states.
Recorded proofLabelled per stageIllustrationExplanationExplanation · sample dataOperator diagramLive read viewMachine-readableAnalysis of public sourcessign-in
Optional advanced review · recorded partial result
Put the governance claims under pressure.
The Governed Escape Test is for teams actively comparing Qensai with an existing or planned agent system. Inspect the declared success conditions, failure boundaries and machine-readable proof contract, then give the reviewed evaluation prompt to an agent you control. It comes after the story, proof and map—not in place of them.