AI and Cybersecurity: Does the Advantage Lie with the Attackers or the Defenders?

At an Operational Risk industry forum this week, an interesting question was raised: is artificial intelligence more beneficial to those attacking our systems, or those defending them? There was no clear consensus. AI is transforming both sides of cybersecurity: increasingly capable frontier models can identify vulnerabilities, write and analyse code, undertake reconnaissance and support complex tasks at remarkable speed. But the same capabilities can help defenders strengthen controls, detect suspicious activity and respond to incidents. So who has the advantage?

The Case for the Attackers

Cyber attackers have always enjoyed an asymmetry: they may need to find only one exploitable weakness, while defenders must protect complex, constantly changing environments. AI could amplify that imbalance.

Highly capable AI systems can search for vulnerabilities, analyse software, generate malicious code and automate elements of the attack process. Increasingly autonomous agents could coordinate activities across multiple systems, increasing speed and scale. Activities once requiring considerable expertise, time and resources may become more accessible through cheaper AI tools. As frontier capabilities diffuse into widely available systems, sophisticated cyber capabilities may reach more malicious actors.

The implications are particularly concerning for financial institutions with complex legacy technology, extensive third-party dependencies and large attack surfaces.

The Case for the Defenders

Defenders have important advantages too: legitimate access to their own systems, operational data, established monitoring arrangements and the ability to redesign environments to make attacks harder. AI can help them:

Identify and remediate vulnerabilities before attackers exploit them.
Analyse large volumes of security information and detect unusual patterns.
Improve threat intelligence and identify emerging attack techniques.
Accelerate incident investigation, containment and recovery.
Continuously test controls and identify weaknesses in security arrangements.

Perhaps most importantly, defenders can use AI proactively to reduce the attack surface rather than simply respond to attacks. The challenge is making effective use of that advantage.

The Critical Differentiator: AI-Powered Detection and Response

One area where AI could fundamentally change the balance is the effectiveness of detective controls. As attacks become faster and more automated, the time available to identify and contain malicious activity may shrink dramatically. Traditional detective controls relying on predefined rules, alerts and human investigation may struggle to keep pace with AI-enabled attacks operating at machine speed.

AI-powered detective controls could become a critical differentiator in determining which organisations successfully defend themselves. AI offers the potential to:

Detect threats earlier by identifying subtle behavioural anomalies and emerging attack patterns.
Improve detection quality by correlating signals across networks, applications, identities and third-party connections.
Reduce false positives and help security teams prioritise genuine threats.
Accelerate investigations by reconstructing potential attack pathways and assessing the scale of a compromise.
Enable faster containment through automated or human-approved defensive actions before an attack spreads.

The key operational risk consideration is the relationship between the speed of an attack and the speed of detection and containment. If an AI-enabled attacker can move across systems in minutes but the organisation takes hours to detect and contain the intrusion, the consequences could be severe. Effective AI-powered monitoring and response could help contain threats before they become significant operational incidents.

However, AI-powered detective controls introduce their own risks. False positives, incorrect automated decisions, compromised monitoring and poorly controlled autonomous responses could disrupt important business services. Speed must be combined with governance, appropriate human oversight and effective control testing.

For operational risk managers, the question is not simply whether detective controls exist and operate as designed, but whether they detect, escalate and support containment quickly enough to prevent unacceptable harm.

The Asymmetry Problem

There is a significant difference between identifying a vulnerability and fixing it safely. An AI agent may discover a weakness in minutes. For a financial institution, remediation may require risk assessment, change approval, testing, coordination with third parties and careful implementation to avoid disrupting critical services. An attacker may need one successful exploit; a defender may need to address hundreds of vulnerabilities.

Could AI therefore accelerate attacks faster than organisations can safely remediate the weaknesses it discovers? That reinforces why preventative and detective controls must work together. No organisation can eliminate every vulnerability, so rapid identification and containment are critical.

Agentic AI Changes the Equation

The emergence of highly capable autonomous agents, and potentially coordinated swarms of agents, adds another dimension. Imagine an attacker using multiple agents to identify weaknesses, analyse infrastructure, develop exploits and coordinate stages of an attack. Now imagine defenders deploying their own agents to monitor systems, investigate suspicious activity, isolate compromised components and support recovery.

We could increasingly see AI systems confronting other AI systems at speeds beyond conventional human decision-making. But autonomy creates risks on both sides. Defensive agents may make incorrect decisions, disrupt legitimate activity, interfere with critical services or introduce new vulnerabilities. The recent Hugging Face incident illustrates how agents in specialised research environments can behave unexpectedly and circumvent intended controls; it does not establish that ordinary enterprise agents will necessarily do the same.

An agent that identifies suspicious activity and recommends action is very different from one authorised to disable systems, revoke access or isolate critical infrastructure without human approval. The appropriate degree of autonomy should reflect the consequences of both the threat and the defensive response.

Measuring the Effectiveness of Detective Controls

The importance of detection and response suggests firms should place greater emphasis on how quickly and effectively controls operate. Relevant indicators include:

Mean Time to Detect (MTTD): How long does it take to identify a potential attack?
Mean Time to Contain (MTTC): How quickly can the organisation contain the threat?
Detection coverage: What proportion of relevant attack techniques and critical systems are monitored effectively?
False-negative rates: How frequently do controls fail to identify genuine threats?
Containment effectiveness: What proportion of incidents are contained before material disruption or harm?

These indicators require careful interpretation. Averages can conceal serious weaknesses, particularly when a few severe incidents take much longer to detect or contain. For operational resilience, the more important question is whether detection and containment can prevent disruption from exceeding the firm's impact tolerances.

A detective control that eventually identifies an attack may be functioning technically, but may not be effective enough to protect an important business service. This is the distinction between a control operating as designed and a control being effective in managing risk. Are second-line assessments testing detective controls against yesterday's threat speeds or tomorrow's AI-enabled attacks?

What Does This Mean for Financial Services?

1. The speed of risk management

Can vulnerability management, patching and change management respond quickly enough to AI-accelerated threats? Are firms using AI appropriately to identify and prioritise weaknesses?

2. Operational resilience

Are firms prepared for faster-moving incidents spreading across systems and important business services? Have severe but plausible scenarios been updated to reflect AI-enabled attacks?

3. Third-party risk

Are critical technology providers investing sufficiently in AI-enabled defence? How do they govern their own autonomous agents? What assurance can they provide about detection, response and recovery?

4. AI governance

What controls govern defensive AI agents, particularly those able to change systems or security configurations autonomously? How are their decisions monitored, challenged and overridden?

5. Assurance and oversight

How can the second and third lines provide meaningful challenge and assurance over increasingly sophisticated AI-enabled cybersecurity? These considerations belong within existing operational risk, cybersecurity, operational resilience and TPRM frameworks, not a disconnected AI governance exercise.

The JADEtc Academy Perspective

Perhaps AI does not inherently favour attackers or defenders. It amplifies capability: potentially greater speed, scale and sophistication for attackers, and better detection, stronger controls and fewer vulnerabilities for defenders. The benefits will not be distributed equally.

Organisations with mature cybersecurity, effective governance, strong operational risk management and the ability to adopt AI safely may strengthen their defensive advantage. Those with weak controls, legacy technology, poor visibility of dependencies or inadequate investment could become increasingly vulnerable.

Critically, the quality and speed of detective controls — increasingly powered by AI — may determine whether an attempted attack becomes a contained incident or a major operational disruption.

The real risk may not be that AI makes attackers inherently more capable than defenders, but that attackers adopt AI faster and more effectively than the organisations they target.

For Boards and risk managers: Are our defensive capabilities — particularly our ability to detect, investigate and contain threats — evolving as quickly as the attacks they are intended to prevent?