Key Takeaways

• The Transport for London (TfL) and Marks & Spencer (M&S) attacks demonstrate the growing importance of identity as the primary attack surface in modern organisations.

• Public reporting suggests both incidents relied heavily on social engineering and compromised identities rather than novel software vulnerabilities.

• The principles described in NIST SP 800-207: Zero Trust Architecture assume that identities, devices and networks may all become compromised and are designed to reduce the impact of successful intrusions.

• Zero Trust should not be viewed as a technology or product. It is an architectural approach that continuously evaluates trust, limits privileges and assumes that attackers may already be present within the environment.

• While no security architecture can guarantee prevention, Zero Trust provides a framework for reducing attacker freedom, limiting lateral movement and improving organisational resilience.

  ---

Artificial intelligence has rapidly become one of the most influential technologies in modern cybersecurity. Over the past few years, security teams have embraced AI to improve threat detection, accelerate investigations and reduce the burden placed on analysts working within increasingly complex SOCs. Large language models can summarise incident reports, generate search queries, explain malware behaviour and help analysts navigate enormous volumes of telemetry. Machine learning systems identify behavioural anomalies that might otherwise go unnoticed, while autonomous AI agents are beginning to automate repetitive investigative tasks.

These developments are transforming security operations. Analysts spend less time writing complex queries or manually correlating logs from dozens of different platforms and more time interpreting evidence, making decisions and understanding attacker behaviour. AI has not removed the need for skilled cybersecurity professionals, but it has made their expertise considerably more scalable.

Continuous Agentic Pentesting for the Modern Web

The only enterprise-grade agentic penetration testing platform built for modern web applications. No consultants. No delays. Up to 10x testing capacity at the same cost of manual.

Learn more

As organisations become increasingly dependent upon AI, however, a new challenge emerges. Security teams must not only use AI to defend their organisations; they must also ensure that the AI systems themselves are secure. An insecure AI model can become an attack vector just as readily as an unpatched server or misconfigured cloud workload.

This shift has led to growing interest in frameworks that apply established cybersecurity principles to artificial intelligence. One of the most prominent is Google’s Secure AI Framework (SAIF), which provides guidance for building, deploying and operating AI systems securely throughout their lifecycle. Rather than introducing entirely new security concepts, SAIF adapts familiar practices such as defence in depth, Zero Trust and continuous monitoring to address the unique challenges presented by machine learning and generative AI.

For security professionals already using AI to enhance their operations, understanding SAIF offers a valuable next step.

AI is changing how security teams work

Security operations have traditionally been constrained by three persistent challenges: an overwhelming volume of security data, a shortage of experienced analysts and the increasing sophistication of modern cyber attacks.

A typical enterprise may generate millions of events every day from endpoints, firewalls, identity providers, cloud services and network monitoring platforms. Although Security Information and Event Management (SIEM) platforms have helped centralise this information, analysts have still been required to search, filter and correlate vast quantities of telemetry before meaningful investigations can begin.

Artificial intelligence significantly reduces this workload. Modern security platforms increasingly allow analysts to interact with systems using natural language rather than platform-specific query languages. Instead of remembering complex search syntax, an analyst can simply ask for endpoints exhibiting suspicious PowerShell activity combined with unusual authentication behaviour during the past week. AI translates that request into appropriate search queries, retrieves the relevant data and often provides an initial interpretation of the results.

Large language models are also proving valuable during investigations themselves. They can explain unfamiliar command-line arguments, decode obfuscated scripts, summarise lengthy log files and map attacker behaviour to frameworks such as MITRE ATT&CK. Tasks that previously required extensive manual research can now be completed within seconds, allowing analysts to devote more attention to validating evidence and understanding the broader context of an attack.

Machine learning extends these capabilities further by identifying behavioural anomalies across users, devices and cloud environments. Rather than relying exclusively on signatures or known indicators of compromise, AI identifies subtle relationships between seemingly unrelated events. An isolated authentication anomaly may appear harmless, but when combined with unusual process execution, privilege escalation and unexpected cloud activity, it begins to resemble a coordinated intrusion.

This ability to correlate weak signals across diverse sources is one of AI’s greatest contributions to modern security operations.

AI as a force multiplier rather than a replacement

Despite concerns that artificial intelligence may eventually replace cybersecurity professionals, the reality within most organisations is considerably more nuanced. Today’s AI systems function most effectively as force multipliers.

They automate repetitive analytical tasks, accelerate investigations and reduce the time required to move from an alert to a meaningful conclusion. They do not eliminate the need for human judgement. Experienced analysts still determine whether suspicious behaviour represents genuine malicious activity, assess organisational risk and decide how incidents should be handled. Business context remains difficult for AI to infer. Administrative maintenance, software deployments or emergency infrastructure changes frequently resemble attacker behaviour until organisational knowledge is applied.

This partnership allows security teams to investigate more incidents without proportionally increasing staffing levels while also reducing analyst fatigue and improving investigation quality.

The challenge of securing AI

As organisations integrate AI into critical business processes, AI systems themselves become attractive targets. Traditional software security focuses on protecting applications, operating systems, networks and data. AI introduces additional attack surfaces.

Large language models can be manipulated through prompt injection attacks designed to override intended instructions. Training data may be poisoned before models are deployed, influencing future decisions in subtle ways. Sensitive information can potentially be extracted through model inversion attacks, while attackers may attempt to steal proprietary models that represent significant intellectual property.

Agentic AI introduces additional considerations. An autonomous AI system connected to enterprise infrastructure may possess permission to query databases, access cloud environments or interact with operational systems. If those permissions are excessive or poorly governed, compromising the agent may provide attackers with extensive access across the organisation.

These risks do not invalidate AI adoption, of course. They simply demonstrate that AI systems require the same disciplined approach to security that organisations already apply to traditional technology.

Introducing Google’s Secure AI Framework

Google developed the SAIF to help organisations address these emerging challenges.

Rather than presenting AI security as an entirely new discipline, SAIF builds upon decades of established cybersecurity practice. Principles such as least privilege, defence in depth, secure software development and continuous monitoring remain highly relevant. The difference is that these principles must now be applied to models, datasets, prompts, autonomous agents and machine learning pipelines as well as conventional applications.

This practical approach makes SAIF particularly valuable for security professionals. Most organisations do not need to abandon existing security frameworks when adopting AI. Instead, they extend familiar controls into environments where intelligent systems increasingly participate in operational decision making.

The framework encourages organisations to consider security throughout the AI lifecycle rather than treating it as a final deployment activity.

Models should be designed securely, training data should be protected, inference environments should be monitored and operational behaviour should be continuously evaluated. This philosophy closely mirrors the evolution of DevSecOps, where security becomes embedded throughout software development rather than added after applications have already been built.

Applying familiar security principles to AI

One of SAIF’s greatest strengths is that it avoids reinventing cybersecurity. Instead, it demonstrates how established security practices naturally extend into AI environments.

Strong identity management remains essential. AI services should authenticate users securely and operate with the minimum permissions necessary to perform their functions. Zero Trust principles continue to apply. AI agents should not automatically trust requests simply because they originate from internal systems. Every interaction should be authenticated, authorised and validated.

Defence in depth remains equally important. Organisations should combine multiple protective controls rather than relying upon individual safeguards. Secure infrastructure, protected APIs, encrypted communications, access controls, behavioural monitoring and comprehensive logging collectively provide resilience against evolving threats.

Continuous monitoring also becomes increasingly important. Unlike traditional software, AI systems may change in effectiveness over time as user behaviour evolves, models encounter unfamiliar inputs or attackers develop new techniques for influencing outputs.

Security therefore becomes an ongoing operational activity rather than a deployment milestone.

AI makes security operations easier

Perhaps the most immediate benefit of AI lies in how much easier it makes daily security work. Analysts frequently spend large portions of their day performing relatively routine tasks: interpreting log entries, translating encoded commands, documenting investigations, reviewing historical incidents and searching multiple systems for related evidence.

AI dramatically reduces this administrative burden:

• Investigation summaries can be generated automatically.

• Threat intelligence reports can be condensed into concise operational guidance.

• Detection rules can be converted between different query languages.

• Security documentation becomes easier to produce and maintain.

• Junior analysts can receive explanations of unfamiliar attack techniques without interrupting senior colleagues.

• Routine activities such as preparing incident reports become significantly faster.

As these efficiencies accumulate, security teams gain additional capacity to perform higher-value work including proactive threat hunting, detection engineering, purple team exercises and strategic security improvements. The result is not merely faster investigations but a better use of human expertise.

Building confidence in AI-assisted decisions

Although AI offers considerable advantages, security teams must remain confident that automated recommendations are trustworthy. Transparency therefore becomes essential. If an AI system recommends isolating a production server or disabling a privileged account, analysts should understand why.

Explainable AI allows investigators to review the evidence supporting recommendations rather than accepting opaque conclusions. Modern security platforms increasingly provide attack graphs, evidence timelines and references to the underlying telemetry used during analysis.

This visibility encourages analyst confidence while also supporting audit requirements, regulatory compliance and forensic investigations. SAIF reinforces this philosophy by encouraging organisations to validate AI outputs continuously rather than assuming models will always behave correctly. Human oversight remains an essential safeguard.

Preparing for the next generation of security operations

The evolution of AI within cybersecurity is still in its early stages.

Today’s AI assistants are already reducing analyst workload through natural language interfaces, behavioural analytics and automated investigation support. Tomorrow’s autonomous AI agents are likely to coordinate increasingly complex investigative workflows, gathering evidence from multiple platforms, consulting threat intelligence and recommending response actions with minimal human intervention.

uture analysts will spend less time performing repetitive data analysis and more time directing AI systems, validating conclusions and making informed security decisions. Understanding how AI operates, where it can fail and how frameworks such as SAIF help secure these technologies will become an increasingly valuable professional skill.

Looking ahead

Artificial intelligence is rapidly becoming part of the operational fabric of modern cybersecurity. It enables security teams to investigate incidents more quickly, understand complex attacks more easily and manage environments that have grown beyond the limits of manual analysis. By automating repetitive work and enhancing behavioural detection, AI allows organisations to strengthen their security operations without requiring proportionally larger teams.

As AI becomes embedded within enterprise infrastructure, however, it must be protected with the same discipline applied to every other critical technology. Google’s Secure AI Framework provides a practical roadmap for achieving this by extending established cybersecurity principles into the world of machine learning, large language models and autonomous AI systems.

The future of cybersecurity is therefore unlikely to involve humans competing with intelligent machines. Instead, it will involve security professionals working alongside AI systems that enhance their capabilities while remaining secure, trustworthy and carefully governed. Organisations that embrace both sides of this equation—using AI to improve security operations while securing AI itself—will be better prepared for the increasingly complex threat landscape ahead.

Protecting your Identity with a Zero Trust Mindset Austin Miller · April 1, 2025 Read full story

#243: The Entrepreneurial Adversary Austin Miller · Jul 17 Read full story

#244: Comparing Two Modern UK Cyber Attacks Austin Miller · Jul 24 Read full story