Pluto / Anthropic reconciliation — 2026-11-12

Scope: Anthropic-related sources discovered through Pluto’s blog inventory and Plutonium guides, including its linked Claude Code repository. This is a supplemental source-to-control review, not a live tenant assessment. The existing Anthropic coverage ledger retains the wider November 5 product census, Trust Center artifacts and federal/commercial distinctions. Stable controls were extended rather than renumbered.

Source dispositions

Each row covers recommendation families rather than reproducing source text. Covered means the repository contains guidance, not that a customer’s deployment passes. Rejected means not adopted as a universal/native control, not that every underlying risk is false.

Pluto source Recommendation families and disposition
Securing Managed Agents Covered by API 3.4: restricted egress, default-denied tools, approval for powerful actions, narrow mounts, application authorization for custom tools and advisory instructions. Added: precise vault header-only substitution and workspace authorization, package/network coupling, per-thread monitoring, lifecycle/export caveats. Shared pack 4.2 now includes a separate native credential fragment. Rejected: stale environment shape, package installation with package-manager access disabled, and “no dedicated documentation” assertion.
Inside Managed Agents Covered by API 3.4: separate agent/environment/session boundaries, credential proxy, untrusted context, scoped tools and network. Rejected: converting researched internal runtime paths, root/seccomp observations or implementation details into unsupported customer configuration. New vault caveats prevent interpreting placeholder substitution as impossible secret exfiltration.
Cowork hardening Covered by Enterprise 2.1–2.4, 3.3–3.9, 5.1 and Code 5.2–5.3: access/role rollout, browser domains, folders, connectors/plugins, network/search, logging, secret hygiene, updates and response. Rejected: blanket audit exclusion, desktop MCP-file governance of all surfaces, local MCP inherently safer, and generic endpoint encryption as a new Anthropic-specific control. Current cloud/local and retention distinctions remain authoritative.
Inside Cowork Covered by Enterprise 3.3, 3.6–3.8, 5.1: browser outside VM, local evidence, remote initiation and persistent work. Rejected: undocumented ports/flags/internal architecture as supported admin switches. No generic runtime isolation fix invented.
Extension ecosystem Covered by Code 2.1–2.2, 3.1, 4.1 and Enterprise 3.1, 3.5–3.6: minimize primitives, inspect complete dependency/tool/hook graph, pin trusted sources, distinguish surfaces. Added: skill frontmatter permission preapproval, managed-rule suppression version, and hook exceptions. Rejected: prompts or local placement as enforceable isolation.
Claude Code Best Practices repository README recommendation review, not an audit of every bundled implementation. Covered by Code 1.1–1.2, 2.1–2.2, 4.1–4.2, 5.1, 5.3: central rules, approvals, secret avoidance, telemetry, MCP review, careful configuration merging and recurring review. Rejected: behavioral text as enforcement, an empty MCP file as complete admission policy, command deny rules as a sandbox, full-observability guarantees, local MCP safety assumptions, or a scanner that warns when absent as fail-closed prevention.
Securing Claude Tag Already covered in Tag 1.1–5.2 and existing packs: invoker admission/additive roles, channel authority, narrow credentials versus unioned repositories/plugins, service identities, all egress layers, untrusted prompts, memory, routines/offboarding, spend/DM distinctions, audit and migration. Rejected: turning observed model resistance into an enforcement guarantee or treating server-side search as categorically incapable of disclosing query data.
Inside Claude Tag Already covered by the same Tag controls: service versus human identity, channel scoping, inherited authority, public/private memory, autonomous routines, separate downstream audit, proxy and search boundaries. Current official scope/audit docs were reopened. Rejected: extrapolating successful third-party injection tests to immunity, shared provider IPs to tenant identity, undocumented proxy status to stable API, or historical connector/model counts to current configuration.
Office add-ins Added Enterprise 3.10: independent app admission, connected-account/document scope and synthetic verification of capture. Covered: data crosses Microsoft boundary; per-surface controls cannot be inferred. Rejected as stale: no Compliance capture for these surfaces; current vendor documentation describes M365 capture in beta. Do not assert a universal retention period.
Artifacts publication Added Enterprise 5.3: external sharing, email invitations and individual exceptions are separate; inspect complete artifact source, test non-member access and review legacy/live artifacts independently. Covered: secret read prevention and permission boundaries. Rejected: prompt-only publication approval as a deterministic gate. No undocumented write API or publication inventory invented.
Function hooks Added Code 2.2 and shared pack 7.6: strict profile actually disables hooks; managed/SDK/built-in/force-enabled-plugin exceptions documented. New CLI calls native strict plugin validation without installing or loading the plugin. Covered Code 4.1: source provenance, review executable extensions, update drift, treat terminal credential prompts as untrusted. Validation is not a security verdict; no CVE or vendor fix inferred.
Claude Code vulnerability discussion Covered Code 2–4: hostile repositories, sensitive-file access, runtime/egress boundaries and review extensions. Rejected: generic compromise scenarios as proof every repository-open compromises current clients, unsupported incident statistics or broad “traditional controls fail” claims as a benchmark. No distinct newly documented native setting.
AI-generated code security Covered Code 4.2 and logging/review controls: human accountability, review outputs, dependency governance, secure workflow and evidence. Out of product-control scope: mandatory per-AI-commit SBOM, separate SAST profiles, AI-origin tags and centralized organizational ownership are proposed SDLC policies, not verified Claude settings. Rejected: automatic regeneration or stronger prompts as proof of secure code. Existing repository/CI controls remain applicable regardless of author.
Enterprise Compliance integration Covered Enterprise 2.1–2.4 and existing inventory/export packs: people, roles, resources, changes and review evidence. Not adopted: Pluto’s private-preview scoring/classification/graph claims as native Anthropic capabilities or a requirement to buy Pluto. No additional Anthropic enforcement interface established.
Introducing ClaudeSec Discovery source for the guide index and extension capability review, already covered Code 4.1 / Enterprise 3.5–3.6. Launch counts, ratings and connector anecdotes are not authoritative vendor defaults or current package security verdicts. No subscriptions, external contact, or automatic installation of the referenced tooling.

Current first-party evidence / six-surface census

Retrieved November 12 unless explicitly marked prior evidence. This closes implementation selection for these incremental changes, not a claim that every provider resource was freshly enumerated.

Surface Evidence and decision
Terraform Community registry returned a JavaScript shell this pass. Earlier full census in November 5 ledger remains prior evidence, not fresh validation. No new Terraform capability inferred; these gaps are local policy, credential configuration and manual admin validation, not falsely modeled provider resources.
REST API Vaults, secure deployment, environments, event list. Native beta shapes checked before the credential fragment was added. No request sent to a tenant. Header/body placement, allowed hosts and workspace caller authority remain independent.
CLI Plugin reference and plugin security: native strict validation; MCP checks require 2.1.281+. Official ant CLI reviewed for API surface, not confused with local Claude Code. The new script wraps only supported local validation; no model/plugin execution.
SDK Official Python SDK and vendor vault/event examples show native beta resources. Existing SDK packs retained. No custom SDK or invented wrapper required for these additions.
Configuration Settings reference, skills, Managed Agent sources above. Chosen surfaces are managed JSON policy and separate API configuration fragments. Skill suppression is explicitly version-gated at 2.1.282+; hooks and plugins retain stated exceptions.
Logging / SIEM Monitoring, events, Compliance capture, Tag audit. Existing collectors retained; no fictional SIEM table or comprehensive-capture guarantee. Verify actual downstream receipt, beta feature scope and thread coverage.

Additional primary checks: artifact admin controls, Tag scope inheritance, Cowork organization controls. These supersede older third-party defaults and exclusions.

Trust Center continuity and access limitation

Purpose-built Trust Center discovery was not skipped. The Anthropic catalog failed this pass through web retrieval and a bounded direct request (TLS timeout). The prior ledger preserves the successfully retrieved catalog, eight-page Cowork/unified baseline and federal guides with dates/hashes. Those are previously verified, not claimed freshly re-downloaded. Current public first-party configuration pages support the incremental changes here; a fresh Trust Center catalog comparison remains an access limitation. No gated document was bypassed and no account was created.

Validation and boundaries