Cybersecurity in the Age of AI: When Speed Becomes the Threat

Artificial intelligence is changing cybersecurity, but perhaps not in the way many organisations expected.

The most important lesson from this week’s threat landscape is not that autonomous AI has suddenly replaced cybercriminals. It is that attackers are using AI alongside stolen credentials, excessive permissions, compromised developer environments and weak access controls to move faster through systems that were already vulnerable.

For business leaders, that distinction matters.

The challenge is no longer simply keeping attackers out. Organisations must assume that identities, applications and increasingly AI agents may eventually be compromised, and make sure that one compromised account cannot become the key to the entire business.

AI Is Accelerating Existing Attack Paths

Recent AI-assisted intrusions demonstrate how conventional security weaknesses can become considerably more dangerous when combined with automated tooling.

Reported ARTEX-related activity affected organisations including WorkOS, Chainstack and Chainalysis, while CrowdStrike reported the framework being used against more than 200 organisations. Importantly, the reporting still identified human involvement in post-compromise activity. This points towards a hybrid threat model in which attackers use automation to accelerate reconnaissance, credential exploitation and movement through compromised environments rather than relying on completely autonomous attacks.

The lesson for businesses is simple: AI does not need to invent a new vulnerability to increase cyber risk. It can make existing vulnerabilities easier and faster to exploit.

A forgotten API key, excessive service-account permissions or poorly protected administrative interface becomes substantially more valuable when an attacker can automate elements of discovery and exploitation.

When Developer Access Becomes Business-Critical Access

This risk becomes particularly serious around development environments.

This week’s reporting highlighted an extraordinary example involving Binance, where attackers reportedly compromised engineers through malicious developer tools and fake recruitment assessments before abusing stolen credentials, an internal AI tool and wallet-signing infrastructure.

Regardless of industry, the broader control lesson is significant.

An engineer may need extensive access to build, test and deploy applications. But development credentials should not automatically provide a pathway to financial approvals, production secrets or other high-impact business systems.

Organisations should be asking:

• Are developer and production privileges properly separated?
• How long do privileged credentials remain valid?
• Can an AI agent execute actions using a human user’s authority?
• Which actions require independent approval?
• Can compromised credentials be revoked quickly across every connected system?

AI makes these questions more urgent because agents can perform actions at machine speed.

Your AI Agent Is Now an Identity

As organisations deploy coding assistants and autonomous or semi-autonomous agents, security teams need to stop treating them merely as software features.

They should increasingly be treated as digital identities.

This week’s intelligence reinforces why. Researchers identified thousands of exposed secrets associated with Claude Code configurations and repository histories, while other research demonstrated how browser-extension functionality could potentially bridge a browser session with desktop command execution. Stolen enterprise AI credentials are also reportedly being monetised through so-called “LLM mixers”, creating potentially substantial fraudulent usage charges.

Giving an AI agent persistent credentials and broad permissions creates many of the same risks as giving an employee permanent administrator access, except that the agent may be capable of executing actions much faster.

Strong AI governance therefore needs technical controls behind it.

Businesses adopting AI agents should consider:

Separate identities.
Agents should operate through specifically assigned identities rather than inheriting unrestricted employee credentials.

Short-lived credentials.
Tokens and secrets should expire quickly wherever operationally practical.

Least privilege.
Agents should receive only the permissions needed for defined tasks.

Independent approval.
Financial transactions, production changes, credential modifications and other consequential actions should require controls outside the AI model itself.

Comprehensive logging.
Organisations need visibility into what an agent accessed, what actions it attempted and what ultimately changed.

MFA Alone Is Not the Finish Line

Another important development this week involves phishing campaigns using device-code authentication and adversary-in-the-middle techniques to steal corporate sessions and tokens.

These attacks illustrate an uncomfortable reality: conventional multifactor authentication does not stop every form of credential or session theft.

For organisations already using MFA, the next step is to strengthen how authentication itself is designed.

Phishing-resistant methods such as FIDO2 security keys and passkeys can provide stronger protection than authentication methods that still allow users to unknowingly approve attacker-controlled login flows. Organisations should also review whether device-code authentication is genuinely required and restrict it where there is no legitimate operational need.

Cyber Risk Is Also Becoming Physical Risk

Perhaps the clearest reminder that cybersecurity is no longer confined to IT came from attacks targeting industrial refrigeration systems.

Reported incidents across Europe affected supermarket refrigeration controllers, with attackers forcing systems into defrost cycles or unsafe temperatures. Consequences reportedly included food losses, equipment damage and fires, while an attempted attack also reached blood-storage infrastructure.

For Australian retailers, healthcare organisations, manufacturers and logistics providers, this is a useful warning.

Internet-connected operational technology needs to be treated as part of the organisation’s cyber attack surface.

Default credentials, unsupported firmware, contractor remote-access pathways and internet-exposed controllers can ultimately create physical, financial and safety consequences, not merely data-loss risks.

What Australian Businesses Should Do Now

This week’s threat landscape points towards five practical priorities:

• Inventory machine and AI identities alongside human accounts.
• Remove standing privileged access wherever short-lived access can be used instead.
• Separate developer, production and high-impact transaction authority.
• Adopt phishing-resistant authentication for administrators and other high-risk users.
• Include operational technology and third-party remote access in cyber risk assessments, particularly where compromise could affect physical operations.

These measures complement established approaches such as the ACSC Essential Eight rather than replacing them.

The Bigger Picture

AI will continue to increase the speed and scale of cybersecurity activity on both sides.

But organisations do not need to predict exactly how sophisticated the next generation of AI-powered attacks will become.

They need to control what happens after something is compromised.

The businesses best positioned for this environment will be those where identities have limited authority, credentials expire quickly, critical actions require independent verification, operational systems are segmented, and unusual activity can be detected and contained before it becomes a business-wide event.

The future of cyber resilience will not depend on trusting AI to make the right decision. It will depend on designing systems where neither a person, a compromised account nor an AI agent has unlimited power to make the wrong one.