AI Security Funding Adds Evidence to the Business of Controlling What Agents Can Do

This is an evidence update to our earlier bank-agent controls Signal, not a first discovery of agent-permission governance. A newly disclosed attack adds fresh evidence to that existing thesis. Earlier financing and a deployed managed-security service provide supporting context.

The next expensive cybersecurity problem may not be an AI system giving a wrong answer.

It may be an AI system taking the wrong action, at machine speed.

On 11 November, Reuters reported that cybersecurity researchers at CloudSEK had uncovered an espionage operation involving autonomous AI tools.

The researchers linked the activity to a suspected Kazakhstan-connected group and said more than 2,000 security-related documents had been stolen from Tajik government systems.

According to the researchers, the attackers used an open-source agent framework called Hermes to plan parts of the operation, create tools and organise stolen information. CloudSEK reportedly discovered the operation through an exposed attacker server.

There are important limits.

The underlying activity began months earlier, not on 11 November.

The fresh development is the research disclosure.

The attribution and document count are CloudSEK's findings, not independently confirmed facts from the affected governments.

The same day, Singapore's Cyber Security Agency published its November Microsoft security-patch alert. Among the critical entries were an Entra Device Registration vulnerability rated 10.0 and an Azure OpenAI privilege-related vulnerability rated 9.8.

These scores describe the severity of reported vulnerabilities; they do not establish that every user has been attacked or that the vulnerabilities caused the Tajik incident.

Now follow the money.

As older context, a cybersecurity company called Daylight announced US$80 million in Series B funding, led by Goldman Sachs Alternatives, with Okta Ventures and other investors participating. The Wall Street Journal first published its funding report on 10 November at 12:00 UTC, or 20:00 SGT, 33 minutes before this edition’s evidence window began.

Its service combines AI-powered security operations with human specialists.

The company reported revenue growth of more than 40 times over the preceding year, although that figure was not independently audited in the retrieved sources.

An Israeli technology report was published at 14:39 local time on 10 November, approximately 20:39 Singapore time. That later coverage falls inside our 24-hour window, but it does not make the previously reported financing a new in-window announcement.

The amount is an announced financing round, not a forecast of future venture funding. It is retained as older supporting context, not counted as a fresh event in this edition.

Our database sees three different pieces coming together: attackers are delegating actions to AI; businesses have to manage severe software and identity risks; and investors are financing systems that combine automation with specialist human judgement.

We think the most valuable security service could become one that proves what an AI agent was allowed to do, what it actually did, and how quickly it could be stopped.

AI-agent permission audits could examine which systems an organisation's agents can reach, what information they can access and whether sensitive actions require independent approval.

Continuous monitoring of agent actions could help organisations notice unusual data access, unexpected tool use or repeated attempts to carry out prohibited work.

The service would need reliable records rather than merely an AI-generated summary.

Managed containment and recovery could provide a specialist team to suspend compromised access, preserve evidence, restore safe operations and help the customer understand the incident.

These activities are not entirely new.

Daylight already describes agent-risk detection deployed for Miro through its managed security service.

This is an existing competitor and commercial deployment claim, not evidence of an empty market or independently verified customer profitability.

The proposed commercial improvement is to combine them around autonomous agent behaviour, rather than treating each AI tool as an ordinary employee account or waiting until an alert becomes an incident.

Listed businesses to watch include Okta (NASDAQ: OKTA), whose venture arm joined Daylight's funding round, CrowdStrike (NASDAQ: CRWD) and Palo Alto Networks (NASDAQ: PANW). Their products overlap parts of the wider security problem, but the evidence does not establish that any of their shares are undervalued or that they will capture this specific new demand.

The practical first move is a controlled test on a customer's own systems. Ask a company to list its AI agents, connected accounts, permitted actions and emergency shutdown methods. Then test whether it can detect and stop an unauthorised action without exposing customer information.

Our forecast: Over the next 30 to 180 days, security buyers may become more interested in contracts covering agent permissions, incident records and human approval.

The earliest proof will be paid evaluations, not the number of alarming articles about AI hackers.

Report this passage