Encased — AI agent security

Customer-hosted · independent review candidate

Encased broker 0.3.2

Keep provider credentials and permitted effects outside untrusted Python. The broker runs on your dedicated host, receives bounded requests from a credentialless, network-none worker, checks the workspace decision and executes the frozen registered operation itself.

Candidate, not production-qualified containment. Installing this package does not isolate an agent. The Linux builder run is recorded below; independent review and tests for your deployment are still required. The website does not run these host tests for you.

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.

Registered actions—not blanket approval

The supplied host example registers a fixed document read and, only when configured, a fixed-model call. “Registered” means available to request; each call still needs the applicable Encased decision. Human approval does not add a new tool or override transport and budget limits.

kb.read · default worker action

Read one fixed HTTPS document

The host selects the document URL and explicitly approves its domain and global IPs/CIDRs. The worker calls client.call("kb.read", {}); it cannot choose a destination or supply request content.

The supplied transport validates every DNS answer, pins the TLS connection and refuses redirects and ambient proxies. The default worker makes no model call, although host and service usage may still cost money.

llm.chat · optional host configuration

Call a fixed OpenAI chat model

The host owns the provider key, model, current input/output prices, input bound, output cap and persistent token/spend ledger. The worker submits 1–20 text messages with user or assistant roles, not system messages, tools, images, a model, endpoint, credentials or retry policy.

Requires explicit approved addresses for api.openai.com and host-only configuration. Calls are non-streaming with automatic retries disabled. Reservations precede dispatch and are not refunded. A real provider call requires your key and credit; no such qualification is claimed.

Not registered by this release

No arbitrary URL fetch, general proxy, host shell, financial transaction or destructive tool is included in the reference broker. Subprocesses inside the restricted worker are not individually approved. Additional customer tools require typed validation, reviewed effect transports, failure handling and their own tests; dashboard policy examples do not create executable broker handlers.

What checks surround a call?

  1. The broker validates the registered tool and arguments and claims the request ID in its host-private journal.
  2. Workspace policy decides allow, hold or deny. A held call waits for a matching human decision; invalid replies, denial and deadlines prevent dispatch.
  3. Acknowledged intent evidence and applicable budget reservation precede the effect. The broker—not the worker—executes the frozen action.
  4. Completion is recorded. An uncertain outcome requires reconciliation, not blind replay; a new request ID is not business-level deduplication.

Host supervision applies worker CPU, memory, PID and elapsed-time limits. The broker cannot reliably detect prompt injection or semantic drift, undo remote effects, or turn dashboard key revocation into a running-process kill switch.

How to test the candidate

01 · VERIFY THE ARTIFACT

Use a disposable host and synthetic data

Use a dedicated Linux VM without customer data or production credentials. Check the package fingerprint before extraction. Run the local suite in a virtual environment and retain its complete output; skipped Docker probes are not a pass.

curl -fL https://encased.ai/sdk/encased-0.3.2.tgz -o encased-0.3.2.tgz
printf '%s  %s\n' 'fc0461a37ef92309be9e18ffb4f26a7332e415db0de0db490d6454a069c8be3f' 'encased-0.3.2.tgz' | sha256sum -c -
# Continue only if the fingerprint check reports OK.
tar xzf encased-0.3.2.tgz
cd encased-0.3.2
python3 -m venv .venv
. .venv/bin/activate
pip install '.[http,providers]'
set -o pipefail
python -m unittest discover -s tests -v 2>&1 | tee local-tests.log

Package SHA-256: fc0461a37ef92309be9e18ffb4f26a7332e415db0de0db490d6454a069c8be3f

02 · CLOSE QUALIFICATION BLOCKERS

Review scripts and pin real images

Release 0.3.2 corrects the two test-code defects found in 0.3.1; the reviewer should still check the probes. Reviewed base and worker image fingerprints remain unresolved and must come from the host run. Changed source needs a newly packaged artifact and fingerprint; never label a modified run as the unchanged release.

Review the base, build the worker, publish it to your private registry and pull its actual manifest digest. A mutable tag or local image configuration ID is not a substitute. Use the script inside the extracted package, not a standalone downloaded copy: it relies on adjacent source, Dockerfile and tests.

03 · RUN THE REAL CONTAINER PROBES

Run on the reviewed Linux Docker host

After resolving the blockers, the packaged script checks the source build against the reviewed worker image and runs both container suites. Inspect blocked TCP, UDP, IPv6, DNS and metadata routes, detached children, credential absence, mounts, resource limits and forced worker removal. Capture host-side traffic and cgroup observations as well as test output.

# Run inside the extracted package on the reviewed disposable Linux host.
# These must be actual reviewed repository@sha256 manifest digests.
read -rp "Reviewed base image digest: " REVIEWED_PYTHON_BASE_DIGEST
read -rp "Reviewed, pulled worker image digest: " REVIEWED_WORKER_DIGEST
export ENCASED_DOCKER_TEST_IMAGE="$REVIEWED_WORKER_DIGEST"
set -o pipefail
bash deployment/run-linux-isolation-tests.sh "$REVIEWED_PYTHON_BASE_DIGEST" \
  2>&1 | tee linux-isolation.log

These are the shipped-script commands, not a claim of a completed run. A zero exit status alone does not close the known probe defects or establish independent assurance.

04 · VERIFY DECISIONS AND RETAIN EVIDENCE

Exercise a disposable workspace end to end

Keep the workspace key on the host only. Start with a fixed synthetic document, then exercise allow, review, approval, denial, timeout, gateway failure and revocation against the real broker. Match request IDs, broker dispatch evidence and dashboard read-back. Optional provider testing needs an explicitly authorised, capped paid run.

Give the reviewer untouched logs, UTC run time, OS/kernel and Docker versions, CPU/memory, source revision, package fingerprint, actual image manifest digests and reviewed configuration with secrets removed. Remediate findings and retest before accepting any deployment-specific containment claim.

Recorded evidence

Local tests passed
73 / 82
Docker probes skipped
9 · not passes
Independent conclusion
Pending

Historical local Python 0.3.2 evidence dated 22 October 2026, separate from the Linux builder run and third-party findings. Local tests and synthetic provider responses do not establish actual container containment or certification.

Source, limits and complementary tools