Published November 10, 2026 | Version report-2026-11

Bastion: Execution Boundaries, Control Evolution, and Bounded Evaluation of a Governed Agent Fleet

Authors/Creators

  • 1. Independent researcher, Stockholm

Description

Bastion is a deployed, three-agent fleet that reviews Google Cloud Identity and Access Management (IAM) access. It was built so that deterministic code, rather than a language model, holds every policy decision. An Access Auditor reads IAM under a read-only identity and gives the model only opaque finding identifiers, departments, reason codes and bounded scores. Model-free steps apply human exceptions and thresholds, route findings to owning departments and require delivery receipts. A separate Escalation Agent, which cannot read IAM, creates human-review records.

This report is an evidence-grounded account of the fleet in four parts: how it was designed, where its execution boundaries failed, how its controls evolved in response, and what the final bounded evaluation did and did not demonstrate. Between August and November 2026, five dated failures exposed gaps between local execution signals and downstream work. A completed status, a worker-local state key, a state-only policy decision, an error-free remote reply or a closed response stream did not establish that the intended work had run across the boundary. Repairs made policy execution, handoffs and downstream completion explicitly checkable.

The final bounded evaluation covered G and H on v0.9.0 and I–M on v0.9.1 on 7–8 November 2026 under pre-specified protocols. One cohort recorded 98 of 98 validated serial production investigations before an instrument stop. Two earlier cohorts are retained as stopped: one on an instrument fault and one on a model-service rejection that failed closed. Targeted controls passed: a direct completion check, five completed-event replays, a prompt-injection refusal and the rejection of a forged routing envelope. A read-only telemetry census found Runtime graph spans in 5 of 142 agent traces. A timeout fault injection observed fail-closed handling, redelivery and dead-lettering. It also observed worker computation continuing after the caller had timed out. The evaluation supports bounded observations on one project and one workload. It is not a reliability rate, a service-level objective, a comparative privacy result or a general security guarantee.

Files: bastion-technical-report-2026-11.pdf (the report, 25 pages); source-manifest.json (SHA-256 of every cited source, script, section, figure, dependency file and licence notice); source-snapshot.zip (those source bytes). The same files are attached to the GitHub release report-2026-11 (tag on commit e9e86b0f46e2cc1ba50897de7dff4f3104117b95).

SHA-256:
bastion-technical-report-2026-11.pdf  ec710124910bba89dcb4fd3bbd769f714347074f3989869c00b74f31a013dcb3
source-manifest.json  b8606936efff176be3b89f2de77fa6fec59095530ff235ad66c5fb6aebbf891a
source-snapshot.zip  3433623337a5069130fbfb1f545fdc1ef4cca6a4325fce588d5ad59e1515a912

Licences: the report text, figures and evidence records are licensed under CC BY 4.0; the code in the source snapshot is licensed under MIT. LICENSE-CONTENT.md in the snapshot gives the scope of each.

Files

bastion-technical-report-2026-11.pdf

Files (1.4 MB)

Name Size
md5:be976675a46b35f4d965f289880a4129
741.7 kB Preview Download
md5:ab849954cded71fbcee3d663bd7bbde1
8.3 kB Preview Download
md5:3c4fa22d69032997d957fbdb40bec631
630.5 kB Preview Download

Additional details

Related works

Software

Repository URL
https://github.com/iarjunganesh/bastion
Programming language
Python