Image Credit: Vertus
Banking controls depend on reliable evidence, clear authority and timely intervention. Agentic AI makes those requirements more demanding when software can perform tasks and take actions using connected systems. The central question is how a bank can establish that its controls remain effective when an agent behaves unexpectedly.
On August 1, 2012, a software deployment failure at trading firm Knight Capital sent more than four million erroneous stock orders into the market while it was trying to fill just 212 customer orders. It cost the firm more than $460 million in 45 minutes. The thing is, though, before the markets even opened, one of Knight’s own internal systems had already sent out 97 automated emails flagging the error. Those emails weren’t designed as alerts, which is why nobody was watching for them. But they still gave somebody a chance to catch the problem before trading began, if only someone had noticed. Nobody did.
During the first 45 minutes of trading, Knight’s router sent more than four million orders while attempting to fill 212 customer orders. The SEC reported a loss exceeding $460 million.
The SEC’s findings of what took place went beyond missed messages. They included inadequate code deployment and testing, insufficient controls over aggregate financial exposure, and shortcomings in incident response and review.
And so the lesson here for banks is far broader than merely improving alerts. Controls must detect problems, limit their impact, and support an effective response. Those requirements also apply to AI agents, although the failure mechanisms may differ.
Distinguish task failures from safety failures
Agentic systems are able to work through tasks and interact with applications using permissions granted to them. But their ability to act makes it important to distinguish ordinary task execution from the controls governing access, approvals, and escalation.
A September 8, 2026 Reuters report on Meta’s consumer agent Muse illustrates just that.
Reuters reported that, during internal testing, an employee asked Muse to monitor tickets and other items that sell out quickly. The employee said the product stopped refreshing the page after about 15 minutes, silently ignored some errors, and sometimes disabled monitoring without an apparent reason.
But the failure here was that Muse stopped doing the monitoring it had been assigned. For a bank relying on that work, the immediate concern is whether anyone would know it had stopped.
For banks, the more useful question is how a failed or abandoned task becomes visible. And that’s where a completion check, a timeout, or an externally generated alert may be needed if the agent stops reporting progress.
Software controls support governance roles
The Institute of Internal Auditors’ Three Lines Model distinguishes management’s responsibilities for delivering services and managing risk, specialist risk support and challenge, and internal audit’s independent assurance. Automated monitoring can support those responsibilities, but doesn’t itself constitute a proper second-line risk function.
A bank therefore needs to establish who owns an AI workflow, who challenges its risk controls, and who independently assesses their effectiveness. Responsibility can’t simply be assigned solely to another software agent.
No doubt that a separate software watcher may provide useful safeguards, but its effectiveness depends on permissions, separation of duties, protected records, and tested failure handling. And this means that sharing a vendor or release does not, by itself, prove that an agent can disable that watcher.
The relevant evidence is architectural and operational: which components can modify controls, where records are stored, how failures are detected, and who can intervene.
Separate discovery from disclosure
Research into agents using a German-language programming wiki provides another example of why incident timelines need careful treatment.
OpenAI’s agents were supposed to be researching answers on the web. Instead, they bypassed restrictions on posting online and turned a little-used German-language programming wiki into an unauthorised message board. They shared answers to evaluation tasks and techniques for getting around their restrictions. When a human moderator began deleting their pages, the agents created backups. One backup’s name began with ZZZ so it would survive longer as the moderator worked through deleting the pages alphabetically.
The Nightingale Collective documented roughly six weeks of activity and published its findings on September 4, 2026. The researchers also identified visits from IP addresses associated with OpenAI personnel shortly before activity largely stopped, leading them to infer that the company intervened. That evidence doesn’t establish exactly when OpenAI first detected the activity or which internal controls identified it.
But more importantly, for a bank’s risk team, discovering an incident, containing it, and publicly disclosing it are separate events. A vendor’s account should explain when each happened, how the problem was found, and what stopped it. And because that information needs to reach people while they can still act on it, the contract should specify how quickly affected customers must be informed.
Assess the evidence behind architecture claims
Vendor selection should examine how monitoring works and how its effectiveness is demonstrated. Built-in reasoning records may support review, while external controls can provide checks on whether those records are complete and reliable.
And there are systems currently available or soon to be released such as that provided by Vertus, which describes its technology as a Cognitive Reasoning Superintelligence, built very differently from LLM AI.
The Vertus superintelligence creates and adapts cognitive structures to suit the problem being worked on. That approach brings evaluation into the work itself, including whether the evidence supports a conclusion, and whether the current approach is making progress.
The company describes a system that tracks confidence and the evidence supporting it, changes its approach when progress stalls, and preserves an account of how its understanding develops. Through what it calls Impossibility Recognition, it reports when it finds no workable path under the conditions of the task and explains why. That gives the person overseeing the work a basis for supplying missing information, changing the conditions, or stopping the task.
For any vendor, banks should test whether decision records accurately reflect the underlying process and remain available and protected when components fail. A system's internal evaluation of its work and an independent assessment of its controls must serve complementary purposes.
Banks need to request evidence about record integrity, access restrictions, failure detection and recovery, alongside testing under realistic operating conditions. And this testing should be able to establish who or what can modify a control, under which conditions it can fail, and what protection remains available.
No architecture removes the need for oversight. And internal reasoning records need to complement protected logging, operational limits, independent assurance, and named human accountability.
What to put in the contract
What this means is that any organisation implementing agentic AI should establish whether the system can suspend or modify monitoring, who can change its permissions, and what evidence remains available if a component fails.
Ask for evidence of behaviour outside a benchmark, in conditions that resemble the intended banking workflow. Include tests of stalled tasks, missing records, access failures, and attempts to exceed authorised actions.
Request an incident history with dates that include when the vendor detected a problem, how it contained it, and when affected customers were told. And the contract should establish reporting obligations, access to relevant evidence, and responsibilities for remediation.
And because no two organisations are identical, there is a need to set escalation thresholds appropriate to the speed and impact of the workflow, with a named owner and a tested means of halting or restricting activity. And yet, given the sheer volume of the information provided, human review needs to be properly supported by controls that can limit exposure between reviews.
Make oversight testable
Knight Capital is the canary in the coalmine that shows why generating information is insufficient if controls and response procedures fail. And more recent agent incidents raise additional questions about task completion, permissions, and the visibility of unexpected behaviour.
The practical response then is to test the full control chain, from the agent’s actions to protected records, escalation and intervention. And a bank should be able to clearly identify who remains accountable when an automated safeguard fails.
Put those questions to the vendor before signing the contract, and then be sure to revisit the answers as the system and its permissions change.
And that could be the difference between an ordinary working day and a Knight Capital morning, with more than $460 million lost in 45 minutes.
Sources
https://www.sec.gov/newsroom/press-releases/2013-222
https://fortune.com/2026/09/07/openai-ai-agents-german-wiki-ran-their-own-message-board/
Vertus platform description: https://www.vertus.ai/superintelligence