Skip to main content
All resources
AI Governance, Ethics & Compliance
AI-assisted · human reviewed before publication

How Can Your Firm Keep Approved AI Plugins From Installing Different Code?

An approved plugin can install code that differs from what was reviewed. Check vendor fixes, test with non-client fixtures, and document the limits of each approval.

|7 min read
AI GovernancePlugin SecurityVendor Due DiligenceLaw Firm Operations

Consider a hypothetical situation: your technology provider approves a plugin for an AI coding assistant. Later, that plugin updates while staff are working. Who checks that the installed code is still the code someone reviewed? As of November 2026, the Plugin4Shell disclosure makes that a concrete operating question for law firms using coding assistants internally or through implementation partners.

The short answer: we recommend making approval cover the installed code and the update process, alongside the plugin’s name. Identify affected tools, verify the vendor’s fix, and test approved plugins with invented files and dummy credentials before allowing access to firm systems. These are proposed operating controls, not evidence that a particular firm has been compromised.

What the disclosure establishes—and what it does not

Air Security’s Plugin4Shell researcher disclosure describes a failure in how Claude Code, Codex, GitHub Copilot, and Gemini CLI handled plugins pinned to a particular code revision. This guide uses that researcher disclosure, an Anthropic advisory, and an OpenAI proposed code change as public technical evidence. It contains no AlphaForge deployment results or first-party law-firm measurements. The supplied source excerpts were accessed on November 14, 2026; their publication dates were not supplied.

The disclosure describes reviewing plugin code at a particular Git commit—a recorded code revision—and pinning installation to that commit’s SHA identifier. Pinning means requesting that reviewed revision rather than whatever code happens to be newest.

Air Security reports that an attacker controlling the plugin repository could make the installation process select different code while the requested pin still appeared intact. The missing safeguard was a check that the actual installed revision matched the requested revision. The researcher describes a route through background plugin updates that required no further user interaction.

Our interpretation is that this disclosure concerns software supply-chain integrity: the route from a reviewed component to the code that actually runs. An assistant’s useful answers might tell the firm nothing about whether its plugin installation process preserves the approved revision.

The disclosure does not establish that your firm installed an affected version, experienced an attack, or exposed client information. It also supplies no law-firm adoption measurements or evidence of lost inquiries or revenue. A conclusion about your firm would require its own records and evidence from its providers.

Separate affected behavior from a verified fixed installation

Anthropic’s Plugin Commit SHA Pin Bypass advisory identifies the affected installation behavior in Claude Code and Claude Desktop. It says the process used Git checkout without verifying that the resulting HEAD—the revision actually selected—matched the pinned SHA. Its described exploitation conditions include attacker control of the upstream repository and a Git host permitting branch names shaped like commit hashes.

The published CVE record identifies Claude Code versions before 2.1.179 as affected and identifies 2.1.179 as the release that fixes this issue. For manual installations, we recommend confirming the installed version with the person responsible for the tool.

OpenAI’s Verify Git plugin SHA checkouts pull request describes checking the resolved revision after checkout and rejecting a SHA-pinned source when it differs from the requested SHA. It also describes a regression test using a default branch named after another commit’s hash. The partially loaded excerpt does not establish merge status, a released fix, or the first fixed Codex version.

The researcher names Copilot and Gemini CLI, but their vendor advisories and release boundaries are not included in the supplied evidence. Their remediation status therefore remains unconfirmed here.

We recommend recording the exact installed version, update channel, vendor advisory or release reference, first fixed version when confirmed, and the person who verified the installation. Mark missing affected-version and fixed-version boundaries as unconfirmed. Keep a proposed code change separate from evidence of a released fix and a verified installation.

Build an inventory around access and responsibility

Consider this invented operational example: an outside developer uses a coding assistant to maintain a firm’s website integration. The firm approved the integration work, but nobody recorded which plugins the developer uses. This is a hypothetical scenario, not a reported client incident.

That arrangement may leave the firm without enough information to assess a plugin advisory. The inventory could address the assistant’s access to firm work, including access through the contractor.

We recommend documenting each relevant installation with the following information:

  • Record the installation. Identify the assistant, installed version, device or working environment, and responsible owner.
  • Record the plugin approval. Identify each plugin’s source repository, approved revision, and update settings.
  • Record accessible resources. List the firm folders, repositories, credentials, and connected services the installation can access.
  • We recommend assigning responsibility. Name the person responsible for reviewing advisories, approving updates, and withdrawing access.

Check both staff-managed and provider-managed tools that touch firm systems. Ask providers for evidence specific to the environment they use for your work, including the installed version and the reference used to confirm remediation.

Test approved plugins with non-client fixtures

For these recommended tests, prepare invented files and expected results as non-client test fixtures. We recommend using fictional matter names, synthetic documents, dummy credentials, and a separate test repository. Keep real client files, production connections, and working secrets outside the test environment.

For the hypothetical website integration, the fixture might contain a fictional inquiry and a test destination that cannot send messages to clients. We recommend using it to observe installation, updates, and permitted behavior without involving a live matter.

We recommend asking a qualified technical owner to perform and document these acceptance checks:

  1. Confirm the installed revision. For a SHA-pinned plugin, compare the revision actually installed with the approved hash. Record how the comparison was made.
  2. Check mismatch rejection. In an isolated environment, use a controlled repository to test whether a requested pin that resolves to different code is rejected. Do not test against a live third-party repository without authorization.
  3. Exercise the update path. Test installation and updating separately. Record whether updates occur automatically and whether the verification check still applies after an update.
  4. Check the agreed access boundary. Run the plugin’s intended task against the fixture. Record file access, command execution, external destinations, and any request for additional permissions.
  5. Rehearse withdrawal. Disable the plugin, remove its access, and document how to return to the approved configuration.

These are recommended acceptance tests; AlphaForge has not performed them on the named products for this guide. Passing them could provide evidence about the tested version and configuration. It would not establish that every plugin behavior or future update is safe.

Renew approval when relevant conditions change

We recommend attaching approval to a recorded configuration: assistant version, plugin source, approved revision, permissions, update behavior, and test evidence. Define which changes require another review, including a new repository owner, changed permissions, or a different update mechanism.

Consider temporarily restricting plugin access to firm systems when the technical owner cannot establish remediation status. Choose that restriction according to the actual access and work involved. This is an operating recommendation, not a conclusion that every named product must be removed.

For the wider decision about access and human oversight, review what your firm should verify before an AI agent handles client work. We recommend keeping plugin verification within that broader approval review.

Measure evidence gaps and completed checks

We recommend tracking installations with confirmed versions, plugins with verified revision checks, unresolved remediation boundaries, and completed fixture tests. Record the date and owner of each unresolved item so the next review can distinguish new evidence from an unchanged assumption.

Review failures by type: a mismatched revision accepted, an unexpected update, access beyond the approved scope, or an incomplete withdrawal procedure. Keep attempted tests separate from completed tests, and retain failures alongside successful results.

We recommend asking the responsible person to show which plugin code runs, why that revision is approved, and what happens when it changes. That record could support a reviewable operating decision while making unresolved questions visible to the managing partner.

AI Governance, Ethics & Compliance

Start with the problem in front of you.

The governance review is a focused entry point into the same Firm Intelligence Audit.