Boards are increasingly being asked to approve AI investments that promise faster decisions, lower operating costs and greater automation.
But before giving AI authority, BOARDs should this ONE question:
Who can stop an AI-driven business decision when trust in the system is lost?
Please note we are not talking to switch off the chatbot or terminate a cloud subscription.
Who has the authority, evidence and operational capability to stop an AI-influenced business process before an uncertain output becomes customer harm, financial loss, regulatory exposure or operational disruption?
A quick Board review is expected today to know if an AI system follows the wrong reality.
Last week, Anthropic disclosed that a retrospective review of 141,006 cybersecurity evaluation runs identified three incidents in which Claude models reached the internet from, or while interacting with, a third-party testing environment and then gained unauthorised access to the production infrastructure of three organisations.
In actual sense the models had been instructed that they were operating in a simulation without internet access. Because of a misunderstanding between Anthropic and its evaluation partner, internet connectivity was actually available. The models treated real systems as part of their assigned exercise.
One incident involved access to credentials and a database containing several hundred rows of production data. Another involved the creation of a malicious Python package that was made publicly available and executed by real systems. Anthropic said the affected organisations it successfully contacted had not detected the activity themselves.
Anthropic characterised the events primarily as harness and operational failures rather than evidence that the models had independently developed malicious goals. The models were pursuing the objectives they had been given while operating with an incorrect understanding of their environment.
The deeper lesson in this case is that an AI system may act consistently with its instructions while the organisation’s assumptions about its context, access and boundaries are wrong.
This is where the Organizational Trust will fail.
Anthropic reported that its most recent internal research model stopped when it recognised that the system it was accessing was real. An older model continued in some runs after recognising signs that the environment might be real, rationalising that the real organisation must still be part of the exercise.
So, the question is Should the organisation rely on the AI system to recognise that it should stop itself?
The answer should be NO, Because, the self-correction by AI may lead to rogue payments or industrial operation if that correction lead to a business logic that was not approved by BOARD.
The organisation needs an independent mechanism that can observe, restrict and stop the AI system even when the model itself continues to believe that its actions are appropriate.
In the another very recent case, the security concerned were raise about some inexpensive Android television boxes that allegedly contained AI program capable of rewriting device identities, automating advertisement interactions and routing third-party traffic through owners’ broadband connections. The reported device was using machine vision and other automation to locate advertisements and execute campaigns.
Just imagine your organisation procuring something without being aware of which kind of AI capabilities its coming with. Is it influencing your business logic / extracting data from your organisations and sharing to threat actors.
Automation can exist inside an ecosystem without the asset owner fully understanding its identity, purpose, communications or authority.
The same principle applies to enterprise AI.
A business may procure an AI-enabled SaaS platform without understanding every model, plugin, agent, data source, API, subcontractor or monitoring mechanism supporting it.
The organisation may believe it has purchased a business capability.
In practice, it may have introduced a chain of automated dependencies that can access data, trigger actions and change behaviour beyond the organisation’s direct visibility.
That is why AI Security Governance and third-party resilience cannot be separated.
From 2 August 2026, the European Commission’s AI Office and national authorities are responsible for implementing, supervising and enforcing significant elements of the EU AI Act. The European Commission has stated that the AI Office can request technical documentation, evaluate general-purpose models, require corrective measures and impose fines for non-compliance within its remit.
Transparency obligations under Article 50 also apply from 2 August 2026, including requirements addressing interactions with AI and the marking or labelling of certain AI-generated or manipulated content. A new enforcement team reported by the Associated Press and syndicated by SecurityWeek is intended to support scrutiny of risks including deepfakes, illicit imagery and cyber threats.
The specific obligations affecting an organisation will depend on its jurisdiction, sector, role and use case. Final interpretations should be validated by legal and regulatory counsel.
The organisation may need to demonstrate not only that its AI system performs, but that it can be governed, monitored, interruptible and accountable.
So, lets circle back to our original question who should have stop authority?
Simple Answer is : The board should NOT attempt to operate the AI system. But The board should require these five capabilities
Detect the loss of trust
Management should define observable indicators before deployment.
These may include abnormal output patterns, unexpected data access, unexplained model actions, control-limit breaches, significant drift, excessive override rates, customer complaints or third-party security notifications.
A generic statement that the system will be “continuously monitored” is not enough.
The board should ask what is monitored, against which threshold, by whom and what decision follows a breach.
Stop the AI’s authority
Stopping the user interface may not stop the process.
An AI agent may retain API tokens, service accounts, workflow permissions, queued jobs or access to connected systems. Hence, the CxOs should be able to revoke the AI’s authority independently of the model provider. This may require disabling integrations, rotating credentials, blocking actions, quarantining outputs or moving the system into recommendation-only mode.
A controlled reduction of authority can preserve continuity while limiting harm.
Continue the business without it
Management must know whether the AI-based process can return to human decision-making, a rules-based alternative, a previous model or a reduced service.
This fallback must be documented, staffed and tested. Otherwise, the organisation may recognise that the AI is no longer trustworthy but keep using it because stopping it would also stop revenue or customer service.
That is not resilience. This is AI Slavery.
Monitoring: Preserving evidence
The organisation should be able to review & audit every activity and communication it did. i.e right from what it accessed to what actions it took and which outputs it influenced or any customer transactions were affected.
Logs must be useful for investigation, not merely produced for compliance.
Restore trust deliberately
AI should not return to autonomous operation simply because the immediate technical defect has been corrected. Management should order complete validation of the model, its data, access, integrations, monitoring and decision boundaries.
The business owner—not only the technical team—should accept the residual risk.
If customer-impacting decisions are involved, the organisation should also review prior outputs.
The real test is not about proving an AI system will ever be wrong. The real test is whether the organisation can recognise that trust has been lost before the impact leads to material loss.
So, would you apply a CyberResilience Strategies in AI Era to establish Control over AI’s Authority to act and Business to continue and also identify decisions that were wrongly influenced by AI.
Because, AI resilience is not the ability to keep AI running at all costs.
It is the ability to stop AI safely, continue the business responsibly and restore trust deliberately.
(Authored By Ajay C Bhayani- Cyber Resilience Strategist, Director AmbiSure Technologies Pvt Ltd.)
