draft: add MCP sandbox authority boundary profile

README.md

@@ -0,0 +1,74 @@
+# MCP Sandbox Authority Boundary Profile
+
+**Draft v0.1 - private review artifact**
+
+A vendor-neutral profile for evaluating whether an MCP-connected execution sandbox
+keeps its declared authority boundary across repositories, storage, credentials,
+network access, lifecycle changes, and evidence outputs.
+
+## Why this profile exists
+
+A consolidated interface can reduce tool-schema and token overhead while
+concentrating more capability behind a single invocation surface. A sandbox can
+provide execution containment without, by itself, proving credential scope,
+resource isolation, egress policy, or revocation behavior.
+
+This profile records **what a target actually declares and demonstrates**. It
+makes no claim about any provider's implementation or security posture.
+
+## Scope and safety
+
+Run only against an explicitly authorized, isolated target with disposable
+identities and synthetic data. Do not use production credentials, customer data,
+or unrestricted egress. Each result must be classified as `simulation`,
+`staging`, or `authorized_live_characterization`; results do not generalize
+beyond that environment.
+
+The profile is intentionally not an exploit kit. It defines boundary assertions,
+required evidence, and stop conditions. Target-specific procedures belong in an
+approved test plan.
+
+## Profile modules
+
+| Module | Question |
+| --- | --- |
+| Resource scope | Can the workload reach only explicitly attached repositories and storage? |
+| Credential scope | Are workload credentials absent by default, narrowly scoped, and revocable? |
+| Egress scope | Is outbound traffic denied or restricted according to an explicit policy? |
+| Lifecycle | Does attach, detach, expiry, and revocation promptly change effective authority? |
+| Evidence | Can an independent reviewer determine the asserted action, policy, and outcome? |
+
+The machine-readable profile is at `profile/sandbox-boundary-profile-v0.1.json`.
+An evidence-record shape is at `schemas/observation.schema.json`.
+
+## Synergy with existing work
+
+- **Authorized but Composed**: a locally valid action can compose into an
+  aggregate authority problem. This profile makes the resource, identity, and
+  egress parts of that boundary observable per run.
+- **Receipt claim verification**: a signed or format-valid record is not enough.
+  Observations must bind the declared boundary, action, outcome, evidence digest,
+  and the authority that supplied each claim.
+- **Agent Security Harness**: this is a narrow environment-profile companion,
+  not a claim of new harness coverage or a vendor finding.
+
+## Minimal workflow
+
+1. Capture the target's documented policy and configuration as an immutable
+   input digest.
+2. Select only authorized profile cases and define stop conditions.
+3. Run against synthetic fixtures with disposable identities.
+4. Emit one redacted observation record per case.
+5. Validate records locally:
+
+```bash
+python3 scripts/validate_profile.py
+```
+
+6. Publish only the profile and redacted synthetic examples unless written
+   authorization permits target-specific evidence.
+
+## Status
+
+Draft for private review. No Hugging Face product behavior has been asserted or
+characterized by this artifact.

examples/synthetic-observation.json

@@ -0,0 +1,11 @@
+{
+  "profile_version": "0.1-draft",
+  "case_slug": "attached_repository_scope",
+  "classification": "simulation",
+  "policy_digest": "1111111111111111111111111111111111111111111111111111111111111111",
+  "action_digest": "2222222222222222222222222222222222222222222222222222222222222222",
+  "outcome": "denied_as_declared",
+  "evidence_digest": "3333333333333333333333333333333333333333333333333333333333333333",
+  "recorded_at": "2026-07-23T13:30:00Z",
+  "notes": "Synthetic example only. No provider behavior is represented."
+}

profile/sandbox-boundary-profile-v0.1.json

@@ -0,0 +1,76 @@
+{
+  "profile_name": "MCP Sandbox Authority Boundary Profile",
+  "version": "0.1-draft",
+  "classification": "private_review_artifact",
+  "purpose": "Evaluate declared and observed authority boundaries for an MCP-connected execution sandbox.",
+  "authorization_requirements": [
+    "written target authorization",
+    "isolated instrumented tenant",
+    "synthetic data and disposable identities",
+    "named owner, stop conditions, and rollback plan"
+  ],
+  "evidence_rules": [
+    "Bind each observation to a policy/configuration digest.",
+    "Classify evidence as simulation, staging, or authorized_live_characterization.",
+    "Redact credentials, sensitive resource identifiers, and customer data.",
+    "Do not generalize a result beyond the characterized environment."
+  ],
+  "cases": [
+    {
+      "slug": "attached_repository_scope",
+      "module": "resource_scope",
+      "assertion": "Effective repository access matches the explicitly attached scope.",
+      "required_evidence": ["declared_resource_scope", "attempt_class", "outcome", "policy_digest"],
+      "stop_condition": "Access reaches an undeclared resource or evidence contains sensitive data."
+    },
+    {
+      "slug": "attached_storage_scope",
+      "module": "resource_scope",
+      "assertion": "Effective storage access matches the explicitly attached scope.",
+      "required_evidence": ["declared_resource_scope", "attempt_class", "outcome", "policy_digest"],
+      "stop_condition": "Access reaches an undeclared resource or evidence contains sensitive data."
+    },
+    {
+      "slug": "workload_credential_presence_and_scope",
+      "module": "credential_scope",
+      "assertion": "Credential presence, audience, scope, and expiry match the declared workload policy.",
+      "required_evidence": ["declared_identity_policy", "credential_metadata_redacted", "outcome", "policy_digest"],
+      "stop_condition": "A non-disposable credential is encountered."
+    },
+    {
+      "slug": "credential_revocation_effectiveness",
+      "module": "lifecycle",
+      "assertion": "Detach, expiry, or revocation changes effective authority within the declared policy window.",
+      "required_evidence": ["lifecycle_event", "before_after_outcomes", "policy_digest", "timestamps"],
+      "stop_condition": "Revocation impacts a shared or production identity."
+    },
+    {
+      "slug": "network_and_dns_egress_scope",
+      "module": "egress_scope",
+      "assertion": "Network and DNS behavior matches an explicit allow, deny, or proxy policy.",
+      "required_evidence": ["declared_egress_policy", "synthetic_destination_class", "outcome", "policy_digest"],
+      "stop_condition": "Traffic could leave the authorized test destination set."
+    },
+    {
+      "slug": "artifact_log_and_cache_isolation",
+      "module": "resource_scope",
+      "assertion": "Artifacts, logs, and caches do not expose data outside the declared session and resource scope.",
+      "required_evidence": ["declared_retention_scope", "synthetic_marker_result", "outcome", "policy_digest"],
+      "stop_condition": "A real secret or customer record is observed."
+    },
+    {
+      "slug": "direct_interface_policy_consistency",
+      "module": "credential_scope",
+      "assertion": "Equivalent access paths do not bypass the declared authority policy.",
+      "required_evidence": ["declared_policy", "authorized_interface_classes", "outcome", "policy_digest"],
+      "stop_condition": "The procedure would test an unapproved interface."
+    },
+    {
+      "slug": "independently_reviewable_observation",
+      "module": "evidence",
+      "assertion": "An observation binds the claimed boundary, action class, outcome, and evidence digest to their appropriate authorities.",
+      "required_evidence": ["policy_digest", "action_digest", "outcome", "evidence_digest", "classification"],
+      "stop_condition": "Evidence provenance cannot be established."
+    }
+  ]
+}

schemas/observation.schema.json

@@ -0,0 +1,18 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "title": "Sandbox Boundary Observation",
+  "type": "object",
+  "additionalProperties": false,
+  "required": ["profile_version", "case_slug", "classification", "policy_digest", "action_digest", "outcome", "evidence_digest", "recorded_at"],
+  "properties": {
+    "profile_version": {"type": "string", "const": "0.1-draft"},
+    "case_slug": {"type": "string", "pattern": "^[a-z0-9_]+$"},
+    "classification": {"enum": ["simulation", "staging", "authorized_live_characterization"]},
+    "policy_digest": {"type": "string", "pattern": "^[a-f0-9]{64}$"},
+    "action_digest": {"type": "string", "pattern": "^[a-f0-9]{64}$"},
+    "outcome": {"enum": ["allowed_as_declared", "denied_as_declared", "policy_mismatch", "inconclusive", "stopped"]},
+    "evidence_digest": {"type": "string", "pattern": "^[a-f0-9]{64}$"},
+    "recorded_at": {"type": "string", "format": "date-time"},
+    "notes": {"type": "string", "maxLength": 1000}
+  }
+}

scripts/validate_profile.py

@@ -0,0 +1,21 @@
+#!/usr/bin/env python3
+"""Dependency-free structural validation for the draft profile and example."""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[1]
+profile = json.loads((ROOT / "profile" / "sandbox-boundary-profile-v0.1.json").read_text())
+example = json.loads((ROOT / "examples" / "synthetic-observation.json").read_text())
+
+assert profile["version"] == "0.1-draft"
+assert profile["classification"] == "private_review_artifact"
+slugs = [case["slug"] for case in profile["cases"]]
+assert len(slugs) == len(set(slugs))
+assert example["case_slug"] in slugs
+assert example["profile_version"] == profile["version"]
+for field in ("policy_digest", "action_digest", "evidence_digest"):
+    assert len(example[field]) == 64 and all(c in "0123456789abcdef" for c in example[field])
+assert example["classification"] == "simulation"
+print(f"valid: {len(slugs)} boundary cases; synthetic example is structurally consistent")