What's New at Associates AI: November Launch
Agent Pages your Teammates build for your organization, Teammates MCP for reaching them from ChatGPT...
OpenClaw's November 19 advisory describes a restricted role that could change persistent configuration through delegation without approval. Here is how to scope exposure, patch the right release, and test the approval boundary across handoffs.
On November 19, 2026, OpenClaw published a high-severity advisory about restricted roles using delegation: in affected versions, the openclaw tool could apply persistent configuration changes without approval even when policy denied execution and file writes. This is not a report that every installation was compromised. It is a precise warning that a role with limited authority might cross into a more powerful configuration path when handing work off.
That distinction matters to any business putting AI coworkers into routine operations. A support coworker may read a customer ticket and ask another role for help with an account record. A marketing coworker may hand an image task to a specialized role. Delegation makes these handoffs productive. But if the handoff can quietly change which actions are allowed afterward, the approval screen in the original role is not a dependable boundary.
The advisory says the potential consequence is enabling host command execution through a persistent change; actual impact depends on configuration and whether lower-trust input reaches the path. Its first patched stable version is 2026.11.4; extended-stable users can move to 2026.10.33. Before upgrading, OpenClaw recommends setting tools.exec.mode to ask and denying untrusted delegated changes. Treat those as the vendor's stated mitigations, not as proof that a particular deployment is safe without inspecting its effective policy.
This article is not a generic patch notice. It is about the operational mistake the advisory exposes: mistaking permission to request help for permission to alter the rules under which the team works. The fix is to check the running version, the handoff route, the configuration-change boundary, and the evidence that a required approval actually happened.
A useful AI coworker can ask a specialist for help without being able to rewrite its own job description. Those are separate capabilities. The first moves work; the second changes authority. Combining them turns a benign handoff into a possible privilege escalation even when the original role cannot run a shell command or write a file directly.
Imagine an operations coworker assigned to summarize incoming service requests. It may read tickets and ask a scheduling coworker whether a proposed appointment fits the calendar. It may not edit server settings. If the operations role can delegate a configuration request that succeeds without the operator's approval, its direct tool restrictions no longer describe its effective authority. The new setting can persist beyond the current ticket and affect future work, including work triggered by different people.
The point is not that every delegation is dangerous. Delegation is the normal way to divide work among roles. The problem arises when a privileged operation is authorized by looking only at the tool that eventually performs it, or by treating the delegated request as if it came from a trusted operator. A decision made at the destination must retain the original requester's scope, the reason for the handoff, and any approval requirement attached to the requested change.
A task packet might say: “Find the next available slot for this customer.” That text is context. It is not a grant to change account permissions, enable command execution, or disable a review rule. The recipient can use the information to perform its own authorized work; it cannot take the sender's words as an administrator credential.
For every consequential handoff, ask three questions. Who initiated the work, which role is executing this step, and which person or policy authorized the specific side effect? If the system can answer only the middle question, a delegated action may look legitimate in a log while hiding its origin. If it can answer all three but does not enforce the last one before the write, the log is a record of an avoidable failure, not a safeguard.
The November advisory is unusually concrete about this line. It describes restricted roles whose execution and file-write permissions were denied, yet a persistent configuration change could still occur through the openclaw tool. An approval process that guards only the obvious write surfaces misses any alternative route to the same durable state.
Start with the advisory itself, not a headline about “all AI agents.” The relevant facts are the installed release, the restricted role's policy, access to the named delegation path, and the ability to make a persistent configuration change. OpenClaw explicitly limits the advisory to that feature and configuration. It does not revoke the project's trusted-operator model or imply that any authenticated Gateway operator is an attacker.
Check the version of the running installation, not just a manifest, a downloaded package, or an update scheduled for tonight. The advisory lists the first patched stable and extended-stable releases; use its affected-version section to compare your branch and deployment. Record the version alongside the Gateway or agent server identifier. A team with several agent servers needs one row for each installation, because updating one does not update the rest.
Next, find restricted roles with delegation enabled or reachable. A role that cannot directly execute or modify files may seem low risk, but that is the very configuration this notice concerns. Document who can submit work to the role: private chat, group chat, scheduled input, webhook, email, or retrieved documents. Lower-trust material in a ticket or page is data, not authorization. Its presence does not prove exploitation, but it affects how urgently the exposure should be investigated.
Finally, inspect the effective rules that govern persistent configuration changes. Can the role request a change? Can the delegated destination carry it out? Who receives the approval? What does a denial do? Preserve a baseline before changing anything so you can tell an authorized repair from an unexplained pre-patch change. If you do not have a configuration history or approval record, say “unknown”; do not turn missing evidence into an assertion that no unauthorized change occurred.
A compact exposure record is more useful than a red/yellow/green label: running version, release branch, named role, delegation route, denied tools, configuration-write authority, input source, owner, and patch status. It lets a small team decide where to act first without pretending the advisory gives a universal risk score for every deployment.
Bad: A customer-service coworker receives a pasted note in a ticket saying a specialist needs a temporary setting change to finish the case. The role cannot edit files, so an operator assumes it is safe. The coworker delegates a request that changes a persistent server setting; no person sees or approves that specific change. The case closes, but the altered setting remains for the next case. This is a hypothetical scenario illustrating the boundary, not a claim about a known incident.
Good: The coworker sends the case facts to a specialist for read-only investigation. The specialist proposes a specific setting change, explains its business purpose and expected impact, and pauses. A separate approval mechanism evaluates the exact change, initiator, target, and expiry before it can be applied. A denial leaves the setting untouched and routes the case to a named human; an approval produces a traceable decision and a post-change verification. Neither role can manufacture that approval by asking another role to do it.
The difference is structural. A prompt saying “ask first” is not the same as a policy that prevents a write until an authorized person or control plane decides. The latter still works if the coworker misunderstands a ticket, a model changes, or a handoff includes misleading text. For more on where policy should live, see keeping security rules outside the AI coworker's writable workspace.
An upgrade closes the known flaw, but it does not by itself tell you whether a setting changed while the old version was running. It also does not validate the rest of your handoff design. Treat the repair as a short operational sequence with a named owner and evidence at each point.
tools.exec.mode to ask and deny untrusted delegated changes while preparing the upgrade. Keep channel and tool allowlists narrow; disable the affected feature where it is unnecessary. Have the operator apply the settings through their supported configuration process, not through a restricted coworker's own request.These steps are an operator's test plan, not undocumented product commands. Use OpenClaw's security documentation and your deployment's supported configuration workflow to implement them. The advisory is the authority for affected scope and patched versions; your own test proves that the local handoff and approval path behave correctly after repair.
There is a business reason to do the full loop. An AI coworker can work across departments and retain context between runs, so a persistent change is not confined to the task in which it originated. Configuration is part of organizational intent: what a role may do, when it pauses, and who can expand its authority. If that intent can be rewritten through a lower-trust route, the team does not actually control the boundary.
Approvals often look strong on a screen but weak in the surrounding workflow. A person may approve a task summary while the exact requested setting is absent. A specialist may receive a delegated message whose origin is not visible. A system may log a successful tool call but not the policy version that allowed it. None of those records answers the practical question: did the right person authorize this particular durable change before it happened?
The approval object should bind to the target operation, the proposed before-and-after state, the requesting role, the executing role, and a time window. Ideally it also identifies the human who approved and the policy version used. If any of those fields change, require a new decision. Do not treat a generic “yes, help with the customer” as permission to alter server-wide execution rules.
Put the enforcement at the privileged operation. A model's summary can help a reviewer understand the request, but it cannot be the mechanism that grants authority. Likewise, a downstream role's own permission to propose a change does not erase the upstream role's restriction. This is seam design in practice: the handoff carries useful work forward while the authority boundary stays in place.
Small teams do not need an elaborate bureaucracy to make this real. One owner, a short register of consequential settings, a change diff, a narrowly scoped approval, and a negative test after upgrades can cover the first meaningful boundary. As the number of AI coworkers grows, keep the same contract across each new role. A teammate who handles inventory should not acquire finance permissions because a delegated task happened to cross from one channel to another.
The OpenClaw notice is a reminder that patching and design are complementary. Patch the specific bypass promptly; then prove that no route, direct or delegated, can silently rewrite the rules. The objective is not to stop AI coworkers from collaborating. It is to let them collaborate without converting every handoff into a standing administrator credential.
Q: What is the OpenClaw delegation approval bypass?
A: OpenClaw's November 19, 2026 advisory says affected versions could let a restricted role use the openclaw tool to make persistent configuration changes without approval even when execution and file writes were denied. The issue is scoped to the named feature and configuration, not every deployment.
Q: Which OpenClaw versions should I use? A: The advisory identifies 2026.11.4 as the first patched stable version and 2026.10.33 for extended-stable users. Compare your running version and branch to the official advisory, and confirm the running process reflects the upgrade.
Q: Does this mean our AI coworker ran host commands? A: No. The advisory describes a possible path to enabling host execution through a configuration change. Whether that happened in your environment requires evidence from configuration history, approvals, and any relevant action logs; the advisory alone does not establish an incident.
Q: Can we just disable delegation? A: Disable the affected feature if you do not need it, as OpenClaw recommends. If the team relies on delegated handoffs, patch and test them instead. Even with delegation disabled, maintain narrow tool permissions and a separate approval boundary for persistent configuration changes.
Q: Is a human approval message in chat enough? A: Only if the actual privileged operation is bound to that decision and the approved scope is enforced before the write. A conversational acknowledgment or a broad task approval is not an authorization for a later, different configuration change.
Q: What should we test after the patch? A: In isolation, try the same benign configuration request directly and through a restricted role's delegation path. Confirm each is denied or paused for the intended approval, then verify a separately authorized change works and leaves an audit record. Avoid using a dangerous production setting as the test fixture.
If your business needs persistent AI coworkers with clear handoff and approval boundaries, Associates AI Teammates gives you a model-agnostic platform for building a governed team; see how it works and get started.
Written by
Founder, Associates AI
Mike is a self-taught technologist who has spent his career proving that unconventional thinking produces the most powerful solutions. He built Associates AI on the belief that every business — regardless of size — deserves AI that actually works for them: custom-built, fully managed, and getting smarter over time. When he's not building agent systems, he's finding the outside-of-the-box answer to problems that have existed for generations.
More from the blog
Agent Pages your Teammates build for your organization, Teammates MCP for reaching them from ChatGPT...
Agent Pages are web tools your Teammates build for your organization, and the pages talk back: ask a...
Teammates MCP lets you message your Teammates from ChatGPT, Claude, Perplexity or any AI app that su...
Want to go deeper?
Get started today. Hire your first Teammate in minutes and put it to work on what you're reading about.
Get StartedWhat we shipped and what we learned.