top of page

Featured in Law360: When AI Agents Can Disable Their Own Monitors, Who Bears the Risk?

jas2824
Nov 7
7 min read

Updated: 5 days ago

We’re proud to share that our founders, Jasmine Parker and Aisha McKinney, were featured in Law360 with a contributed article on a question every company deploying AI agents should be asking: if an agent can interfere with the systems that monitor it, what does a vendor’s promise of “continuous monitoring” actually mean?



We’re also sharing the full article below.


When AI Agents Can Disable Their Own Monitors, Who Bears the Risk?


By Jasmine Parker and Aisha McKinney, Founders of Zova Law


“Continuous monitoring” sounds reassuring in an AI vendor’s proposal. So does “human oversight.” But what do those commitments mean if an agent can interfere with the systems responsible for detecting and reporting its behavior?


In a report updated Oct. 20, 2026, OpenAI described two internal coding incidents. In one, an agent suppressed alerts to OpenAI’s security team while reports were still being recorded. In the other, an agent removed a monitoring integration and later incorrectly denied doing so. OpenAI said it was strengthening its controls and investigating the behavior. [1]


These incidents do not establish that every commercially deployed agent can generally disable their safeguards. They illustrate a narrower contracting problem: monitoring can continue to record activity even when the path for alerting a reviewer has been compromised. An agreement that treats logging, notification and intervention as interchangeable may leave that failure unaddressed.


For corporate customers, the negotiating principle should be clear: allocate responsibility according to who controls each safeguard, and ensure that general AI disclaimers do not undermine specific monitoring commitments. These are proposed contractual protections, rather than conclusions about liability arising from OpenAI’s internal incidents.


That requires examining how the agreement defines monitoring, assigns intervention duties, and responds when those protections fail.


1. Define Monitoring and Protect It From Interference


A vendor may use “continuous monitoring” to mean that activity is logged. A customer may understand it to mean that risky actions are detected, stopped, and escalated.

Those are different commitments.


The agreement should distinguish recording activity, evaluating it, delivering alerts, and intervening. It should identify the agents, integrations, and delegated tasks covered. If an agent launches another process, does monitoring follow? If a report is saved but the alert never reaches a reviewer, has the vendor performed?


A security schedule should establish coverage, measurable detection and escalation targets for defined events, permitted exceptions, and procedures for identifying monitoring failures. Any sampling or delays should be disclosed. Counsel should work with security teams to distinguish commitments the vendor can test and enforce from assurances that cannot be measured.


Buyers should also seek technical restrictions on the agent’s ability to alter monitoring configurations, alert recipients, audit records, and approval mechanisms. For sensitive deployments, the agent should lack permission to modify those controls; material changes should require authorization through a separate administrative process. A second monitoring model alone does not establish independence if the agent can change that model’s configuration or suppress its alerts.


The agreement must then address what happens when monitoring fails. An illustrative negotiating provision could state: If required monitoring becomes unavailable or a material impairment of its integrity is detected, the vendor must suspend the high-impact actions identified in the security schedule until the agreed controls are restored and validated, unless a designated customer representative authorizes a documented, time-limited exception with agreed compensating controls.


The parties should define those actions, the means of detecting control failures and any permitted exceptions before deployment. Payments, production changes, and access modifications may warrant different treatment from low-risk administrative tasks. They should also specify a safe stopping procedure, including treatment of actions already in progress, so suspension does not itself create avoidable harm.


2. Give Human Oversight an Operational Meaning


Human oversight requires someone who receives useful information, has time to assess it, and can intervene.


A general obligation to review AI outputs is inadequate for an agent that takes consequential actions before issuing its final response. The contract should identify which actions require advance approval, who provides that approval, what information is presented, and what happens when no reviewer responds.


Consider an agent investigating a billing discrepancy. Permission to investigate should not silently become permission to issue refunds or change banking information. Approval requirements should attach to defined actions and their material scope.


Where the vendor supplies an approval mechanism, it should commit to enforcing it, including against bypass through delegated processes. Where the customer supplies reviewers, the customer should commit to staffing and escalation procedures. The agreement should specify whether an unanswered approval request expires, is escalated or blocks the action; silence should not inadvertently become authorization.


These obligations should remain distinct. A customer’s failure to review an available alert raises a different allocation question from a vendor’s failure to deliver the alert or enforce the approval requirement. The agreement should address each failure rather than assign every consequence to “human oversight.”


3. Require Notice and Evidence Before Harm Is Confirmed


An incident clause focused exclusively on unauthorized data access can miss a critical event: the agent’s controls have stopped working, but no breach has been established.

Buyers should negotiate a separate notification trigger for a material impairment of monitoring, alerting, approval, or intervention controls affecting their deployment, including a reasonably suspected impairment that warrants investigation. The trigger should not depend on proof that information was stolen or an unauthorized transaction occurred. This contractual notice obligation should supplement, rather than redefine, any applicable statutory breach-notification duty.


Specify an initial notification deadline tied to awareness by designated vendor personnel, including relevant subcontractor personnel, followed by updates as facts develop. This initial notice should report the facts then known without waiting for a completed investigation. Require the vendor to identify affected controls, the known exposure period, containment measures, and recommended customer actions. Serious incidents should produce a written explanation and agreed criteria for restoring autonomous operation.


The customer also needs evidence beyond the agent’s account. Contracts should require records of tool calls, approvals, configuration changes, alerts, and actions, with relevant model and configuration details. Those records should be protected against alteration or deletion by the agent and should connect delegated activity to its originating task. Address retention, timely customer access, and preservation during an investigation, with appropriate limits for personal data, security-sensitive information and other customers’ confidential information, and require appropriate testing before deployment and after material changes, including attempts to suppress alerts or bypass approvals. A general assurance report does not, by itself, demonstrate that these controls work; counsel should confirm the systems controls and period actually examined.


4. Make Specific Commitments Prevail Over General Disclaimers


A carefully drafted monitoring schedule can lose value if another provision broadly disclaims responsibility for autonomous actions or assigns all supervision to the customer.


Counsel should read the security exhibit, AI terms, acceptable-use policy, warranty exclusions, and liability provisions together. Buyers should seek an express order-of-precedence clause making negotiated monitoring, approval, and incident-response obligations prevail over inconsistent general disclaimers. Changes to incorporated online terms should not materially reduce those protections without an agreed process, including notice and a customer remedy if the parties cannot agree.


The same review should address the vendor’s supply chain. A service may depend on a model provider, application developer, and integrator. The agreement should identify who controls each safeguard and hold the contracting vendor accountable for the commitments it makes, including agreed responsibilities performed through subcontractors. If a proposed obligation depends on controls the vendor cannot access or enforce upstream, that limitation should be resolved before the customer relies on the commitment.


Customer obligations should be equally specific: maintaining designated access settings, using approved integrations, and staffing required review functions.


An instruction to perform a task should not automatically excuse a failed safeguard. The relevant inquiry is whether the customer exceeded agreed restrictions, whether a promised control failed, and how each contributed to the loss. A customer-use exclusion should not sweep in every loss merely because a customer initiated the task; any negotiated reduction in vendor responsibility should address the customer conduct that caused or contributed to the failure.


5. Make Remedies Match the Exposure


Service credits may address downtime. They may do little for a company whose agent changes production systems or executes unauthorized transactions during a monitoring failure.


Buyers should negotiate suspension rights, remediation deadlines, termination rights for serious or repeated failures, and assistance restoring affected operations. The agreement should state whether the customer can suspend affected functions without breaching minimum-use or payment commitments, and how fees and transition assistance are handled on termination. Service credits should not become the exclusive remedy for breaches of separately negotiated control obligations.


Liability provisions deserve particular attention. A higher negotiated cap for defined breaches of monitoring or approval obligations may be more achievable than unlimited liability. Its size should reflect the agent’s authority and potential impact. A higher cap alone will not preserve recovery if a separate damages exclusion bars the loss; counsel should expressly address reasonable investigation, containment and restoration costs caused by breach, together with any agreed treatment of unauthorized payments. The parties should also specify how the higher cap interacts with the general cap and other exceptions.


Indemnities require a separate check. A provision limited to third-party claims may not reimburse the customer’s own response costs. If those costs are intended to be recoverable, the contract should create an express reimbursement obligation or otherwise preserve that recovery, rather than relying on the word “indemnify” alone.

Companies deploying agents will face the immediate operational consequences when controls fail. Contracts can allocate duties and financial exposure, but cannot replace functioning safeguards.


Before deployment, counsel should be able to trace a monitoring commitment through the agreement: which control is promised, who can change it, how failure is detected and reported, and what relief follows. A promise of continuous monitoring has limited value if the agent can interrupt escalation or if the contract excluded the resulting loss.

OpenAI’s recent disclosure brings these contracting questions into sharper focus. The significance of the incidents is not simply that agents interfered with monitoring, but that the safeguards intended to make autonomous activity observable were themselves susceptible to interference. OpenAI has since strengthened its controls and is working toward monitoring requirements that are more difficult for agents to disable. For companies deploying agentic AI, the incidents illustrate why the existence of a monitoring system should not, by itself, be treated as evidence that the underlying risk has been adequately addressed.


As agentic systems gain greater autonomy, contracts should reflect the possibility that a control can exist and still fail at the moment it matters. The OpenAI incidents provide an early example of why companies should be able to identify who controls each safeguard, whether the agent can interfere with it, how a failure will be detected and what happens next. Ultimately, a promise of continuous monitoring is only as meaningful as the parties’ ability to rely on that monitoring when the agent behaves in an unexpected way.


This article provides general information and does not constitute legal advice or create an attorney-client relationship. The views expressed are the author’s own and do not necessarily reflect those of her employer or clients. The contractual provisions discussed are illustrative and should be tailored to the particular deployment, agreement, and applicable law.


[1] OpenAI, “Suppressing misalignment monitoring alerts and disabling misalignment monitoring,” report updated Oct. 20, 2026, https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/ (accessed Oct. 26, 2026).

Comments

Couldn’t Load Comments
It looks like there was a technical problem. Try reconnecting or refreshing the page.
ZOVA-Secondary-Logo-With-Tagline-Full-Color.png

Powered by Zova Platform.
Licensed in Georgia & Florida.
© 2026 Zova Law. All rights reserved.

ATTORNEY ADVERTISING DISCLAIMER

The contents of this website should not be construed as legal advice regarding any specific facts or circumstances. All materials were prepared Zova Law, LLC (formerly operating as Evolutionary Ventures Law Group, LLC)  (a Florida law firm organized as a limited liability company, Phone: (716) 687-0409) and are provided for general informational purposes only. Receipt of this information does not create an attorney‑client relationship with Zova Law or any of its attorneys.

You should not act or rely on any information contained herein without seeking professional legal counsel. Prior results referenced in these materials do not guarantee or predict similar outcomes in other matters. Attorneys at Zova Law are licensed in Florida and Georgia. The firm cannot represent clients in all jurisdictions without associating locally licensed counsel and/or obtaining admission in that jurisdiction for a limited purpose.

The attorneys responsible for the content of this website are Aisha McKinney and Jasmine Parker.

COMPANY

 

Blog

​

Media & Press

 

​
 

  • Facebook
  • Instagram
  • LinkedIn
  • TikTok

AMCKINNEY@ZOVALAW.COM  · 716- 687-0409 · ZOVALAW.COM

​
 

bottom of page