Encased — AI agent security
Containment reference

Trust through evidence · Review brief

Independent security review

Our containment statements need scrutiny beyond our own tests. This is the review scope and evidence checklist for a qualified independent engineer assessing the customer-side isolation controller.

Independent assessment pending. The results and historical assessment below are builder evidence, not an external reviewer’s conclusions. No reviewer endorsement or production containment assurance is presented here.

Discuss an independent review Request my reviewer link

22 October 2026 · Python 0.3.2 · Builder-reported

Current test results

Discovered
82
Passed locally
73
Docker probes skipped
9
Failures reported
0

Skipped probes are not passes. These local results include synthetic transport responses, not cloud-host containment verification. The separate Linux builder run is recorded below; no independent report is available. The download is an evidence summary, not raw output.

Linux Docker probes passed 9/9 on 10 November 2026 (Ubuntu 24.04.5 · Docker 29.1.3 · cgroup v2), with 0 failed and 0 skipped. This is builder evidence, not independent review or production qualification. Independent review and deployment-specific qualification remain pending; this server run does not establish on-device isolation.

View all nine Linux test records

Current package SHA-256: fc0461a37ef92309be9e18ffb4f26a7332e415db0de0db490d6454a069c8be3f

Current observations and open findings

Local functional evidence

Approved-domain and IP transport checks

Fourteen deterministic tests cover domain/IP refusal, private or mixed DNS answers, CIDRs, TLS socket pinning, redirects, proxies and fixed-provider restrictions. Responses are synthetic; these tests do not prove real-container containment.

9/9 Docker probes passed on a builder-run Linux host (10 Nov 2026)

Linux worker boundary

The nine Docker-gated probes, skipped locally, passed on a disposable Ubuntu 24.04.5 host (Docker 29.1.3, cgroup v2, runc with AppArmor and seccomp) using test pack v2. Results archive SHA-256 6d21d0661f605a8cb1a65cf3d6ca4cfdcc45bb5cc1d83579819236f13f2b3294. Builder-executed evidence only: not independently reviewed, not production qualification.

No external report received

Independent findings and retest

No signed independent conclusion or remediation retest is available. Current observations are builder evidence, not reviewer findings or certification.

Deployment qualification pending

Provider budgets and permitted-channel leakage

No billable model run or provider-usage reconciliation was performed. Approved providers can receive permitted message content; network approval does not establish data-loss prevention.

19 October 2026 · Python 0.3.0 · Builder-authored

Historical 0.3.0 builder assessment

Not accepted for production containment

Python 0.3.0 source, broker, runner, typed adapters, deployment instructions and supplied tests. Preview workspace connectivity was exercised using a synthetic host-side action over a real Unix socket; no locked Docker worker or billable model call ran here.

Point-in-time 0.3.0 suite: 66 tests discovered, 57 passed and 9 Docker-gated tests skipped; no failures. These are builder observations, not independent assurance. The current 0.3.2 package adds domain/IP controls and is not assessed by this historical report. No comprehensive penetration test was performed.

Reviewed package SHA-256: 41ebc9ef8d503136024d593e91857a6d54692fb85f5b7c2166918917f9fd071d. This reference applies only to that package; a replacement release requires reassessment.

  1. IR-01 · Independence not established

    Reviewer independence & scope

    Prepared by the AI builder involved in implementation. No external reviewer, signed scope or conflict declaration has been supplied.

    Required next evidence: Appoint a qualified external reviewer and agree version-specific scope.

  2. IR-02 · Source inspected · host untested

    Host & worker boundary

    The runner requests a non-root, credential-free, read-only worker with restricted mounts and privileges. Actual host, image layers, kernel and runtime configuration were not assessed.

    Required next evidence: Inspect the deployed host and pinned image; independently verify provenance and isolation configuration.

  3. IR-03 · Local functional evidence only

    Network & action gate

    Unix-socket dispatch, typed requests, fixed destinations and rejection paths have supplied test coverage. Network denial across native calls, DNS and child processes has not been verified in a real container here.

    Required next evidence: Capture independent host traffic and run all denied-path probes.

  4. IR-04 · Preview connection verified · partial coverage

    Workspace decisions & exact approvals

    A real host broker submitted a synthetic action through the preview gateway; acknowledged completion appeared on the workspace dashboard and the disposable key was revoked. Supplied local tests cover approval and denial, but this is not a complete independently witnessed deployment journey.

    Required next evidence: Repeat allow, quarantine, approval, denial, timeout and revocation from the pinned worker on the target host.

  5. IR-05 · Configured · container measurements pending

    Process gate & stop boundary

    Host-owned resource limits, bounded runtime and container removal are implemented. The nine Docker-gated tests were skipped because no Docker engine is available here; skipped tests are not passes.

    Required next evidence: Measure cgroups, deadline overshoot and detached-child cleanup on Linux.

  6. IR-06 · Ledger tests passed · provider untested

    Token & spend gate

    Local tests exercise persistent conservative reservations, concurrency and overshoot handling. No real-provider billing reconciliation or model-specific tokenizer qualification was performed.

    Required next evidence: Approve a capped provider run and verify current prices, usage bounds and budget reconciliation.

  7. IR-07 · Local replay tests passed · recovery incomplete

    Durability & uncertain effects

    Consumed request IDs survive restart in supplied tests, and unacknowledged telemetry stops dispatch. This is not business-level exactly-once execution; admitted remote effects can finish after stop.

    Required next evidence: Independently inject crash failures and rehearse human reconciliation and downstream idempotency.

  8. IR-08 · Design reviewed · deployment privacy pending

    Secrets, evidence & privacy

    The design keeps keys outside the worker and bounds evidence. Image-layer leakage, deployed access, retention and approved-provider data leakage have not been independently assessed.

    Required next evidence: Review synthetic-marker leakage, data flows and privacy controls on the actual deployment.

  9. IR-09 · Deployment defects corrected · retest pending

    Operations & supply chain

    Qualification previously allowed a mutable base and fabricated a repository digest from a local image ID. Those paths were removed; the script now requires reviewed digests and matching built/pulled image identity. Two refusal regression tests passed, but the full registry/build workflow has not run here. Dependencies use version ranges rather than a fully locked reviewed inventory.

    Required next evidence: Independently retest the registry workflow; pin dependencies and assign patching, backup and incident-response owners.

  10. IR-10 · Release blocked by missing assurance evidence

    Findings, retest & release decision

    No signed external report, independent retest or customer production acceptance exists. A supplied test suite and this assessment cannot substitute for those records.

    Required next evidence: Close findings, obtain a signed external conclusion and record customer risk-owner acceptance.

Evidence status

Builder verification
Functional checks are described in the containment reference. They are self-reported and are not an independent assessment.
Customer-host qualification
Actual Linux container qualification evidence remains required. Supplied tests and skipped probes are not proof that a customer host is secure.
External report
No completed independent report is available on this page. Reviewer identity, date, tested version, findings and retest status must accompany any future assessment.

Who should review — and what they review

Choose an engineer experienced in Linux container isolation, network boundaries, Python security and concurrent durable state. Disclose commercial relationships. If Lucas contributes to the controller, another qualified reviewer must independently assess his work; running our tests alone does not establish independence.

The current target is the Python 0.3.2 isolation candidate: worker image and runner, host broker, exact domain/IP transports, typed adapters, SQLite journals and budgets, plus the real hosted proposal, polling and telemetry paths. The 0.2.x cooperative wrappers are not substitutes for this external isolation boundary. Agree host assumptions and exclusions in writing.

Independent review checklist

Each item needs a recorded result, evidence reference, observations and limitations. Use Pass, Fail, Not tested or Out of scope with a reason. Skipped checks never count as passed.

  1. IR-01

    Reviewer independence & scope

    Record reviewer identity, competence, financial relationships and involvement in building the controls. A contributor cannot independently assess their own implementation. Fix the source revision, package SHA-256, image digests, host configuration and exclusions before testing.

    Evidence required

    Signed scope, conflict disclosure, artifact hashes and environment inventory.

  2. IR-02

    Host & worker boundary

    Inspect the actual Linux host and worker: non-root user, no credentials or runtime sockets, minimum mounts, read-only root, dropped capabilities, no privilege escalation, seccomp and resource constraints. Assess kernel and Docker trust and whether the threat model needs a microVM.

    Evidence required

    Redacted container inspection, image provenance and host hardening assessment.

  3. IR-03

    Network & action gate

    Independently test blocked TCP, UDP, IPv6, DNS and cloud metadata paths, including native calls and child processes. Review the per-run Unix socket, permissions, bounded messages, typed argument validation and fixed destinations; reject arbitrary proxy, URL and shell operations.

    Evidence required

    Host-side traffic observations, denied-path results and broker interface review.

  4. IR-04

    Workspace decisions & exact approvals

    Use a disposable workspace to verify allow, review, approval, denial, timeout, malformed replies, service outage and key revocation. Confirm pending or mismatched decisions never dispatch and the broker executes the frozen approved operation, not worker-supplied replacements. Inspect workspace and reviewer access boundaries.

    Evidence required

    Correlated proposal IDs, authorised review decisions, broker dispatch records and dashboard read-back.

  5. IR-05

    Process gate & stop boundary

    Measure CPU, memory, PID and independent deadline limits. Verify detached and setsid children remain within the worker boundary and container removal terminates its cgroup. Race stop against dispatch and document already-admitted work; do not infer remote cancellation from local termination.

    Evidence required

    Cgroup observations, child cleanup checks, deadline measurements and stop-race results.

  6. IR-06

    Token & spend gate

    Review host-only provider credentials, fixed model and endpoint, explicit current prices, input bound and output cap. Verify reservations precede dispatch, persist across restart, serialize concurrent brokers and are never refunded. Qualify the supported tokenizer bound and measure overshoot with an authorised, capped real-provider run.

    Evidence required

    Ledger reservation records, concurrency tests, pricing basis and redacted provider usage comparison.

  7. IR-07

    Durability & uncertain effects

    Inject failures around claim, admission, dispatch and completion. Verify consumed request IDs cannot replay after restart, missing telemetry acknowledgement prevents new effects and uncertain completion stops the broker. Identify downstream idempotency needs; a new request ID is not business-level deduplication.

    Evidence required

    Crash/restart matrix, journal transitions and documented human reconciliation procedure.

  8. IR-08

    Secrets, evidence & privacy

    Inspect images, worker environment, mounts, logs, journal and telemetry for credentials or sensitive data using synthetic markers. Assess input minimisation, access, retention and deletion. Review approved-channel leakage and redaction limits; do not collect full memory or live customer secrets.

    Evidence required

    Redacted data-flow inventory, secret checks and privacy-owner review.

  9. IR-09

    Operations & supply chain

    Review pinned dependencies and images, installation steps, state-directory permissions, backups, restore, monitoring and bounded logs. Assign patching, credential rotation, incident response and rollback owners; document how administrator state deletion or new runs reset protections.

    Evidence required

    Dependency inventory, recovery rehearsal and named operational responsibilities.

  10. IR-10

    Findings, retest & release decision

    Map every public containment claim to evidence or mark it unsupported. Record severity, impact, affected versions, remediation owner and exclusions. Retest fixes and obtain reviewer conclusions plus customer risk-owner acceptance before any production containment claim.

    Evidence required

    Signed report, findings register, retest evidence and version-specific release decision.

Review package & safe test conditions

Supply exact source and package hashes, reviewed base and worker digests, dependency inventory, redacted host configuration, threat model and prior test outputs. The reviewer should inspect and reproduce the tests on the actual qualified Linux host, then add independent checks. Authorise only disposable workspaces, synthetic data and controlled destinations; cap and approve any billable provider tests separately.

These are review inputs, not certified artifacts. Independently verify that each download matches the agreed source revision before relying on it.

From findings to a release decision

  1. 01 · Report

    A signed report records scope, reviewer independence, methods, artifacts, findings, severity, evidence and untested controls. Keep raw logs, attack details and customer information private.

  2. 02 · Remediate & retest

    Assign each finding an owner. Record fixes against a new revision and have the reviewer retest. Unresolved findings need explicit risk-owner treatment; changing the host or artifacts may invalidate earlier conclusions.

  3. 03 · Bounded acceptance

    Require a reviewer conclusion and customer risk-owner acceptance for the stated environment. Publish a reviewer-approved redacted summary with limitations and report reference, not an unqualified “secure” badge.

An assessment is point-in-time and scope-specific. It cannot establish universal bypass prevention, reverse external effects or create a remote emergency-stop channel. Encased is not independently SOC 2 audited or ISO 27001 certified. The paid customer audit is separate from independent review of Encased's own controller.