What the Anthropic Decision Reveals About the Future of AI Security

The recent decision by the U.S. administration to lift restrictions on Anthropic’s frontier AI models has generated plenty of debate. Some have questioned whether the original restrictions were justified, while others argue they reflected legitimate concerns about the cybersecurity capabilities of increasingly powerful AI systems.

Regardless of where you stand, I believe the real story lies elsewhere.

This is one of the clearest examples yet of governments treating AI models as technologies with potential national security implications rather than simply another software product. That should make every cybersecurity leader take notice.

AI Security Is Different

For decades, cybersecurity has focused on protecting systems from attack. Today, we are entering an era where AI itself can influence the speed, scale and sophistication of those attacks.

Modern frontier models can assist with code analysis, vulnerability discovery, malware understanding and offensive research. While these capabilities also benefit defenders, they raise difficult questions about where the line should be drawn between innovation, responsible access and national security.

Whether governments choose to impose restrictions or not, the conversation itself demonstrates how quickly AI has become a strategic cybersecurity issue.

Don’t Forget the Fundamentals

Amid the discussion surrounding advanced AI capabilities, it is important not to lose sight of the basics.

Cybersecurity expert Dr. Ilia Kolochenko, founder of ImmuniWeb, recently observed that many successful attacks still rely on exposed services, weak identity controls and configuration errors rather than sophisticated zero-day vulnerabilities.

I agree with that assessment.

The vast majority of successful breaches today are not the result of nation-state-level AI capabilities. They occur because organisations fail to manage identities, secure cloud configurations, maintain asset inventories, patch critical systems or validate that their controls are actually working.

AI may change how quickly attackers identify opportunities, but it does not eliminate the need for strong cyber hygiene.

Key Takeaway

AI doesn’t replace cybersecurity fundamentals.

It increases the speed and scale at which both attackers and defenders can operate. Strong identity management, secure configuration, asset visibility and proven governance remain the foundation of effective cyber resilience.

The Real Challenge for CISOs

The lesson from this episode is not whether Anthropic should or should not have faced restrictions.

The lesson is that AI is forcing organisations to think differently about governance.

Security leaders now need to consider questions that barely existed a few years ago:

• Who can access frontier AI models?
• How are prompts, outputs and sensitive data governed?
• What controls exist to prevent misuse?
• How do organisations monitor AI-assisted development and security testing?
• How do we balance productivity with acceptable risk?

These are governance questions as much as they are technical ones.

Security Needs to Stay Evidence-Based

As with any emerging technology, there is a danger of focusing on the headlines rather than the evidence.

Some commentators argue AI is transforming offensive cybersecurity overnight. Others believe the impact is overstated.

The reality almost certainly sits somewhere in the middle.

AI undoubtedly improves the efficiency of many security tasks for both defenders and attackers. However, organisations are still compromised every day through preventable weaknesses that have existed for years.

Security leaders should resist chasing every new headline while neglecting the fundamentals that continue to cause the majority of incidents.

Final Thoughts

The debate surrounding Anthropic’s models is likely to be remembered as one of the early moments when AI moved beyond being viewed solely as a productivity tool and began to be treated as technology with genuine strategic security implications.

Whether future governments choose tighter controls or greater openness, one thing is becoming increasingly clear: AI security is not simply another branch of cybersecurity.

It is rapidly becoming a discipline in its own right.

The organisations that succeed will not be those that react to every headline. They will be the ones that combine strong governance, proven security fundamentals and evidence-based decision making while adapting to a rapidly changing AI landscape.

That is where the real competitive advantage will be found.

When the Frontier Blinks: What the Mythos and Fable Controversy Reveals About AI Security

When Anthropic abruptly pulled Mythos 5 and Fable 5 from circulation, the move sent a jolt through the AI and cybersecurity communities. These were not minor point releases. They were widely regarded as among the most capable models the company had ever shipped, and watching them withdrawn, even temporarily, raised an uncomfortable question: if the frontier itself can be paused over a safety concern, what exactly are we securing, and how would we know if it failed?

At the time of writing, much of the detail remains disputed. Anthropic and government officials appear to hold very different views about how serious the issues really were, and until more technical evidence is made public, nobody outside the organisations directly involved can say with confidence what happened. What we can do is step back and ask why an episode like this matters at all, because the answer says a great deal about where AI security is heading.

A bigger jump than the version number suggests

Part of what made the withdrawal so striking was the sheer capability of the models involved. Each generation of frontier model has tended to do more than simply answer questions a little better. The leap to Mythos 5 and Fable 5 was less about polishing existing skills and more about expanding what the models could attempt in the first place: longer chains of reasoning, the ability to plan and carry out multi-step tasks, smoother use of external tools, and a far greater capacity to operate with limited human supervision.

That shift changes the stakes. A model that can reason its way through a complex problem, call tools, and act over an extended horizon is enormously useful, but it is also a far more interesting target. The more a system can do on its own, the more there is to misuse if its guardrails can be bent, and the harder it becomes to predict every way it might behave. Capability and risk tend to rise together, and these models pushed both higher than their predecessors.

Why AI security is different

Traditional software security is largely about code. Attackers look for a buffer overflow, a missing check, a way to slip past authentication. With AI, the attack surface is often the model itself, and the goal is not to break the software but to influence its behaviour, persuading it to do something it was designed to refuse. That is a fundamentally different problem, and it does not respond to the same defences.

Over the past few years the industry has poured enormous effort into safety controls: guardrails, alignment techniques, content filtering, and monitoring, all intended to stop models from producing harmful output or helping with activities that carry security, safety, or legal risk. The trouble is that attackers are rarely interested in using a system the way it was meant to be used.

Controls only matter under pressure

Every security professional knows this instinctively. Organisations deploy firewalls, endpoint protection, privileged access controls, network segmentation, and monitoring platforms, yet none of these are considered secure simply because they have been switched on. Their real value is proven only when something tries to defeat them. AI models are now entering that same phase.

As capability grows, so does the incentive to probe the edges. Researchers, security professionals, hobbyists, and adversaries alike are experimenting with ways to push models in unexpected directions, through prompt engineering, by chaining interactions together, by hunting for weak points in safety systems, or by testing how different controls interact under unusual conditions. This is not a failure of AI. It is a sign that AI has become important enough to attract serious adversarial attention.

Penetration testing, red teaming, adversary simulation, bug bounty programmes, and vulnerability research all exist for one reason: comfortable assumptions tend to collapse the moment they meet real-world pressure. There is no good argument for treating AI systems as an exception.

From theory to practice

The Mythos and Fable episode shows how quickly the conversation about AI safety is moving out of the seminar room and into the real world. As these systems start to handle money, write and run code, advise on health, and influence decisions that affect ordinary people, the questions stop being abstract and start being urgent:

• How effective are the controls, really?
• How should that effectiveness be measured?
• What level of risk is acceptable, and to whom?
• Who decides when a weakness is serious enough to act on, and who is accountable when it is missed?

These are hard questions, and they touch on more than engineering. They are about public trust, about who carries responsibility when something goes wrong, and about whether the people relying on these systems can see enough to judge them fairly. None of that gets easier as capability climbs.

The real lesson

Whatever the final outcome of this particular dispute, one lesson already seems clear. AI security is no longer mainly about building powerful models. It is increasingly about proving that the controls around those models keep working when capable, motivated people set out to make them fail, and about being honest when they do not.

That is a challenge the cybersecurity profession knows all too well. The rest of the world is about to learn it too.

AI in the UK: Driving Innovation Without Expanding Cyber Risk

Written by Sean Tilley, Senior Sales Director EMEA at 11:11 Systems 

Artificial intelligence is no longer a future ambition for UK organisations. It is already shaping how decisions are made, how services are delivered, and how quickly businesses can respond to change. From automation and analytics to customer engagement and operational optimisation, AI is becoming an integral part of the modern enterprise.

AI Governance and Cyber Resilience: A Boardroom Imperative 

As adoption accelerates, however, a quieter risk is emerging, and it is one that boards and executive teams cannot afford to treat solely as a technical issue. AI is not simply another tool for innovation. It is altering the cyber risk landscape and unsettling long held assumptions about security, governance, and resilience.

 

Recent research by 11:11 Systems highlights the scale of that concern. In a global survey of more than 800 senior IT leaders, nearly three quarters (74%) said they believe integrating AI into their organisations could increase vulnerability to cyber attacks, a view shared particularly strongly by both UK and European respondents. This reflects that while they aren’t reluctant to innovate, there is growing recognition that AI changes how risk behaves, moving faster, spreading more easily and becoming harder for leadership teams to understand  and control.

 

Why Boards should be Paying Attention

AI can strengthen cyber defences. Machine Learning systems are capable of spotting anomalies at speed, automating elements of incident response, and helping security teams prioritise threats more effectively. In theory, these capabilities should favour defenders.

 

In practice, the same techniques are also being adopted by attackers. AI is already being used to generate more convincing phishing campaigns, automate reconnaissance, and adapt malware in real time. UK Government commissioned research has shown that vulnerabilities can arise at every stage of the AI lifecycle, from early design decisions through to deployment and ongoing maintenance This creates new attack surfaces that many organisations are still learning how to manage.

 

For boards, the implication is that AI risk can no longer be contained within IT functions. It raises questions about compliance, reputation, operational continuity, and long-term value, while also challenging how risk is identified, tested, and understood at the board level, particularly when AI-driven systems behave in ways that are opaque or difficult to predict.

 

While the technical risks continue to evolve, two organisational dynamics are making them harder to control.

 

Shadow AI is Becoming Endemic

Employees are increasingly turning to unapproved or unsanctioned AI tools to work faster and more efficiently. Often this happens with good intent, but without visibility, governance, or security oversight. UK regulators have been clear that organisations remain accountable for how personal and sensitive data is handled, regardless of whether AI tools are formally approved or informally adopted.

 

The Information Commissioner’s Office (ICO) has repeatedly emphasised that AI deployments must comply with UK GDPR principles, including transparency, accountability, and data minimisation. When AI use sits outside formal controls, blind spots emerge, making it harder to demonstrate compliance to regulators and auditors and harder to contain incidents when something goes wrong.

For boards, the risk is not simply the existence of unauthorised tools. Fundamentally, the risk lies in the widening gap between what leaders believe is happening inside the organisation versus how AI is being used day to day, under pressure to move faster.

 

Pressure for Speed is Outpacing Resilience Planning

AI initiatives are often driven by competitive urgency. Leadership teams want rapid deployment, visible progress, and quick returns. Yet research suggests this urgency often comes at  the expense of recovery readiness, oversight and confidence in how incidents should be handled. This is supported by the 11:11 Systems study which found that many organisations remain overconfident in their ability to recover from cyber incidents, even as complexity increases.

 

When AI systems are deployed before recovery, backup, and incident response plans have been tested against new threat scenarios, resilience becomes theoretical. In an AI driven incident, the speed and effectiveness of recovery will determine the scale of operational disruption, regulatory scrutiny, and reputational damage the business faces.

 

Why Resilience Models Must Evolve

Many board level approaches to resilience were designed for risks that were visible, testable, and broadly predictable. AI quietly undermines those assumptions.

 

UK organisations are increasingly being encouraged to rethink resilience in light of how AI changes the pace and complexity of incidents. That shift is evident in three areas.

 

Recovery processes are evolving to become more automated and scalable. This reflects the reality that manual responses struggle to keep up with fast moving, complex incidents. Research shows that prolonged recovery times significantly increase financial and operational damage following cyber events, particularly in large enterprises.

 

Testing is changing. Static, annual recovery plans are poorly suited to adaptive threats. Government research into AI security risks points to the need for ongoing validation across the AI lifecycle, rather than periodic, check list driven assurance.

 

Finally, resilience is being treated less as a downstream activity and more as a design principle. Governance, visibility, and recovery capabilities are increasingly expected to be built into AI deployments from the outset, not added after an incident. UK regulatory guidance reinforces the expectation that organisations can demonstrate control and accountability over AI driven processes, even as those systems evolve.

 

The Board Level Takeaway

AI represents a strategic opportunity for UK businesses. But adoption that outpaces governance and recovery planning can quietly expand exposure at the very moment organisations believe they are becoming more advanced.

 

The question for boards is no longer whether to adopt AI, but how to do so responsibly. Confidence in innovation needs to be matched by confidence in recovery. That requires tougher questions about visibility, testing, and readiness, not just performance and productivity.

 

In this context, AI governance is not about controlling technology. It is about restoring board level confidence in how risk is understood and managed. In an increasingly complex UK threat landscape, the organisations that succeed will not be those that move fastest at any cost. They will be the ones that embed cyber resilience into AI adoption from the outset, innovating with intent and remain resilient in the face of increasing complexity.

Controlling AI Agents: Why Detection Is Too Late

This is Part 2 of a 2-part series. Read Part 1: Your AI Agent Doesn’t Care About Your Controls

If AI agents change how execution happens, they also expose a fundamental limitation in how most security controls operate. Many control models assume there is sufficient time to detect, assess, and respond to events before they result in material impact.

That assumption no longer holds.

Traditional detection models follow a sequence in which an event occurs, is logged, analysed, and then acted upon. This approach works at human speed, where actions are spaced out and intervention is possible. In automated environments, particularly those involving AI agents, that sequence is compressed to the point where response often occurs after the outcome has already been realised.

The Speed Problem

AI agents operate at a pace that removes the window for meaningful intervention. They do not pause between actions or wait for approval cycles. Once triggered, they execute tasks rapidly across systems, often chaining multiple actions together in a single flow.

By the time an alert is generated, the action has already completed. By the time it is reviewed, the downstream effects may already be in place. This fundamentally changes the role of detection. It becomes a mechanism for understanding what has happened, rather than preventing it.

The Illusion of Monitoring

In response to increasing complexity, many organisations invest in expanding their monitoring capability. More logs are collected, more alerts are generated, and dashboards become more detailed. However, this increase in visibility does not automatically translate into improved control.

Without context, prioritisation, and validation, monitoring becomes a record of activity rather than a means of assurance. Security operations teams are often left dealing with high volumes of alerts that are difficult to interpret in real time. AI agents amplify this problem by increasing both the frequency and the speed of events.

The result is a growing gap between what can be seen and what can be meaningfully controlled.

Evidence and Its Limitations

Control assurance depends on evidence, but evidence is only useful if it reflects the current state of the environment. In slower operational models, evidence can remain valid for extended periods. In highly automated environments, that validity window shortens significantly.

Access models change, configurations drift, and behaviour evolves rapidly. Evidence that was accurate recently may no longer represent reality. This creates a challenge for organisations that rely on periodic validation or static reporting to demonstrate control effectiveness.

In this context, evidence must be continuously refreshed and validated against actual behaviour, not assumed based on design.

What Needs to Change

Addressing this challenge does not require abandoning existing frameworks, but it does require changing how they are applied.

The first shift is from access control to execution control. It is no longer sufficient to confirm that an identity has access to a system. Organisations must understand what actions are being executed, in what sequence, and under what conditions.

The second shift is the reintroduction of accountability. Every action performed by an AI agent should be attributable, traceable, and explainable. Without that, it is difficult to demonstrate that controls are operating effectively.

The third shift is towards validating real behaviour. Policies, architectures, and intended workflows provide useful context, but they do not prove how systems behave in practice. Validation must be based on observed activity in production environments.

Finally, independent assurance becomes increasingly important. Second-line validation provides a mechanism to challenge assumptions, review evidence, and confirm that controls operate as intended under real conditions. Without this, there is a risk that assurance becomes self-declared rather than evidence-based.

Final Thought

AI agents are efficient and highly capable. They execute exactly what they are designed to do, often with greater speed and consistency than human operators.

The risk lies not in the technology itself, but in the assumption that existing control models still provide meaningful assurance. As execution changes, so must the way control effectiveness is measured and validated.

The key question is no longer whether controls exist, but whether their effectiveness can be demonstrated in the context in which they are now operating.

Your AI Agent Doesn’t Care About Your Controls

This is Part 1 of a 2-part series on AI agents and control assurance.  Read Part 2: Controlling AI Agents: Why Detection Is Too Late

The cybersecurity industry has spent years investing in visibility. Dashboards have improved, detection tooling has matured, and the volume of telemetry available to security teams has increased significantly. Most organisations can now see more of their environment than at any point in the past.

However, one of the most important emerging risks is not hidden malware or an unknown vulnerability. It is the rapid introduction of AI agents operating across environments that organisations do not fully understand, cannot clearly inventory, and often cannot meaningfully govern.

This is not simply another software category. It represents the introduction of autonomous digital actors interacting with identity systems, APIs, SaaS platforms, cloud environments, and business processes. These agents are not constrained by the same assumptions that underpin traditional control models, and that is where the risk begins to surface.

From Users to Actors

Traditional security models are built around users. Users authenticate, request access, and perform actions within defined boundaries. Even when errors occur, those actions are constrained by identity controls, privilege models, monitoring, and the natural pace of human interaction. There is friction in the system, and that friction is part of how control is maintained.

AI agents remove much of that friction. They are not passive tools assisting users; they are active actors executing tasks. They retrieve data, make decisions, invoke APIs, and trigger workflows across multiple systems in seconds. The shift is subtle but important. The challenge is no longer limited to managing access. It becomes a question of controlling execution.

Execution Without Assurance

Most organisations assume their existing control frameworks still apply in this new model. On paper, they do. In practice, they often do not.

Control frameworks were designed to validate human-driven actions, predictable workflows, relatively static privilege models, and slower operational cycles. They were not designed to validate high-frequency automated decisions, cross-system execution chains, or real-time, context-driven behaviour.

This creates a gap that is easy to overlook. The agent may be authenticating correctly, calling approved APIs, and interacting with authorised systems. From a control perspective, nothing appears to be broken. Yet there is often no mechanism to prove that the actions being executed are appropriate, proportionate, or safe in the context in which they occur.

Where Controls Start to Fail

This is not a theoretical issue. It is a structural one, and it tends to appear in consistent ways across environments.

The first area is identity. AI agents commonly operate using service accounts, shared credentials, or delegated access tokens. While this enables integration and automation, it weakens attribution. In a traditional model, actions can be traced to an individual. In an AI-driven model, activity may be technically valid but operationally ambiguous, making it difficult to establish accountability when something goes wrong.

The second area is privilege. To enable capability, agents are often granted broad access across systems and services. However, least privilege is not simply about limiting access; it is about ensuring that access is used appropriately in context. An agent may be authorised to access a system, but that does not mean every action it performs within that system aligns with business intent or risk tolerance. Most control models validate access rights rather than behavioural appropriateness.

The third area is monitoring. As automation increases, so does logging. However, more data does not necessarily lead to more assurance. When an agent executes hundreds of actions in a short period, logs can quickly become noise, alerts become volume-driven, and meaningful signal is harder to extract. Monitoring shifts from proactive oversight to retrospective analysis.

The final and most important area is control validation. Controls such as access reviews, segregation of duties, and approval workflows may still exist, but they are rarely tested against autonomous, multi-step execution across systems. The result is not a lack of controls, but a lack of confidence that those controls are operating effectively in the way they were intended.

Final Thought

AI agents are not bypassing controls. In most cases, they are operating within them. The issue is that those controls were not designed to validate how work is now being executed.

If control effectiveness cannot be demonstrated against real behaviour, then the presence of controls alone does not provide assurance.

Next in the series:
Controlling AI Agents: Why Detection Is Too Late

AI Agents Are Creating a New Cybersecurity Blind Spot

The cybersecurity industry has spent years focusing on visibility. Dashboards expanded. Detection tooling improved. Telemetry volumes exploded. Yet one of the biggest emerging risks in 2026 is not hidden malware or an unknown zero-day. It is the rapid deployment of AI agents that organisations barely understand, cannot fully inventory, and often cannot meaningfully govern.

AI agents are moving beyond chat interfaces and simple copilots. They are increasingly capable of reasoning, planning, accessing systems, invoking tools, retrieving information, and taking autonomous actions with limited human involvement. That changes the security conversation entirely.

This is not simply another software category. It is the emergence of autonomous digital workers operating across identity systems, APIs, SaaS platforms, cloud environments, and business processes.

And most organisations are deploying them faster than they can secure them.

Research and industry reporting throughout 2026 show a growing concern across both government and enterprise sectors around agentic AI security risks. Security leaders increasingly view autonomous AI systems as one of the most significant new attack surfaces facing organisations.

The concern is justified.

AI agents introduce a combination of risks that traditional governance and security models were never designed to handle.

AI Agents Change the Nature of Identity Risk

Most cybersecurity programmes were built around managing human identities and traditional service accounts. AI agents disrupt that model because they behave more like autonomous actors than passive software components.

Many organisations are now deploying AI agents with:

• access to internal documentation
• integration into SaaS platforms
• permissions to execute workflows
• API access to sensitive systems
• delegated authority to make operational decisions

The problem is not simply access. It is scale and autonomy.

Industry forecasts suggest AI agent identities may soon outnumber human identities dramatically inside enterprise environments.

That creates several immediate challenges:

• identity sprawl
• excessive permissions
• unmanaged API tokens
• poor lifecycle governance
• invisible machine-to-machine trust relationships
• difficulty attributing actions and accountability

In many environments, organisations already struggle to maintain accurate inventories of privileged accounts or SaaS integrations. AI agents accelerate that problem significantly.

The result is a growing gap between operational reality and governance visibility.

AI Agents Create a New Attack Surface

The security industry often focuses heavily on model risks such as prompt injection or data poisoning. Those are important, but they are only part of the picture.

The bigger issue is that AI agents operate across interconnected runtime environments.

Modern agents may:

• consume external data
• invoke plugins and APIs
• interact with cloud services
• maintain persistent memory
• chain multiple actions together
• collaborate with other agents
• execute operational workflows automatically

That creates an entirely new form of runtime attack surface.

Recent research highlights risks including:

• memory poisoning
• malicious tool invocation
• indirect prompt injection
• AI supply chain compromise
• runtime dependency manipulation
• self-propagating agent loops
• excessive agency and unauthorised action execution

The important point is this:

Many of these attacks do not exploit traditional software vulnerabilities. They exploit trust, autonomy, orchestration, and context.

That makes detection and governance significantly harder.

Why Existing Security Controls Are Struggling

One of the most dangerous assumptions organisations can make is believing existing security tooling automatically extends to AI agents.

In many cases it does not.

Traditional controls were largely designed for:

• deterministic systems
• predictable workflows
• static permissions
• human-driven actions
• relatively stable software behaviour

AI agents are fundamentally different.

They are probabilistic, adaptive, and capable of unexpected behaviour under changing context conditions.

This creates several assurance problems:

• inventories quickly become outdated
• permissions drift continuously
• actions may not be fully explainable
• logging lacks meaningful context
• governance ownership becomes unclear
• accountability boundaries blur

The challenge is not merely technical. It is operational.

Security teams increasingly face environments where AI functionality appears inside:

• SaaS products
• collaboration platforms
• development tooling
• cloud management interfaces
• workflow automation systems
• productivity platforms

Often these capabilities are enabled by default or adopted informally by business teams before governance frameworks exist.

This is rapidly becoming one of the largest forms of Shadow IT the industry has seen.

The Real Risk Is Governance Lag

The most significant AI security risk in many organisations is not the AI itself.

It is governance lag.

Technology deployment is moving faster than:

• control validation
• identity governance
• operational assurance
• policy adaptation
• board understanding
• security architecture redesign

This creates a dangerous illusion of control.

Dashboards may still appear green while autonomous systems quietly accumulate:

• privileges
• integrations
• external dependencies
• sensitive data access
• operational authority

Without strong governance, organisations risk repeating familiar mistakes:

• deploying first
• governing later
• discovering exposure during incidents

The difference now is speed.

AI systems compress timelines dramatically.

What Security Leaders Should Do Next

The organisations responding most effectively are not trying to ban AI agents entirely. They are focusing on visibility, containment, and evidence-driven governance.

Several priorities are emerging:

1. Build an AI Asset Inventory

Most organisations cannot currently answer:

• which AI agents exist
• what systems they access
• what permissions they hold
• what data they process
• who owns them

That must change quickly.

AI agents should be treated as managed operational assets with clear ownership and lifecycle governance.

2. Apply Least Privilege Aggressively

Many AI deployments currently operate with excessive permissions for convenience.

That is unsustainable.

AI agents should operate with:

• constrained access scopes
• segmented permissions
• time-limited credentials
• monitored API activity
• restricted tool invocation

The principle of least privilege matters even more in autonomous environments.

3. Treat AI Runtime Behaviour as an Assurance Problem

The industry increasingly needs continuous validation rather than static approval models.

Security teams should focus on:

• runtime monitoring
• behavioural drift detection
• evidence freshness
• control verification
• anomalous workflow analysis

This aligns closely with broader Continuous Control Monitoring (CCM) approaches already emerging across cybersecurity assurance programmes.

4. Update Governance Frameworks

Most governance structures were not designed for autonomous operational actors.

Boards, risk committees, and security leadership teams need clearer accountability models around:

• AI deployment ownership
• operational risk tolerance
• human override mechanisms
• auditability
• resilience testing
• third-party AI exposure

The governance gap is becoming as important as the technical gap.

Final Thought

AI agents are not simply another cybersecurity trend. They represent a structural change in how digital systems operate.

The organisations that succeed will not necessarily be those deploying AI fastest.

They will be the organisations that can answer:

• what their AI systems are doing
• what authority they possess
• how they are governed
• how they are monitored
• whether their controls still work under real operational conditions

That is ultimately the real challenge of AI security in 2026.

Not visibility alone.

But provable assurance.

Sources and further reading:

• International AI Safety Report 2026
• State of AI Cybersecurity 2026
• Agentic AI security research papers
• Industry reporting on AI agent attack surfaces and governance risks

Mythos AI: What Security Leaders Should Do Next

The recent discussion around Anthropic’s Claude Mythos Preview and Project Glasswing has caught the attention of the cybersecurity industry for good reason.

Mythos is not just another AI announcement. It is being positioned as a frontier model with advanced cybersecurity capability, particularly around finding and exploiting software vulnerabilities. Anthropic has stated that Project Glasswing is intended to give selected defenders early access to this capability to help secure critical software, rather than releasing the model broadly.

Cisco has also published guidance following its work with Mythos, explaining that it is changing its near-term threat modelling of AI-enabled attackers and issuing defensive recommendations for customers. That is the important point.

Whether Mythos itself remains tightly controlled or not, the direction of travel is clear. AI-enabled vulnerability discovery and exploitation capability is improving quickly. Security teams need to prepare for a world where attackers can find, chain and act on weaknesses faster than many organisations can currently respond.

Why Mythos Matters

The concern is not that every attacker suddenly has access to Mythos today.

The concern is that Mythos shows what is becoming possible.

If AI can accelerate vulnerability discovery, exploit development and attack path analysis, then the defensive timeline changes. Security teams cannot rely on slow review cycles, stale evidence or manual-only response models when the speed of threat discovery is increasing.

This does not mean the fundamentals no longer matter.

It means they matter more.

Cisco’s guidance focuses heavily on strengthening fundamentals such as phishing-resistant MFA, Zero Trust, least privilege for AI agents, disciplined patch management and full asset visibility. It also highlights removing end-of-life systems, automating detection and containment, embedding active defences and using AI defensively for threat hunting, validation and testing.

That is where the practical response needs to start.

The Risk Is Speed

Many organisations still manage cyber risk through processes designed for a slower environment.

• Monthly reporting.
• Quarterly reviews.
• Annual testing.
• Periodic evidence collection.
• Manual triage.
• Long remediation cycles.

Those activities still have a place, but they are not enough on their own.

AI-enabled attackers will not wait for the next governance cycle. They will look for exposed systems, weak identity controls, unpatched vulnerabilities, misconfigured cloud services and overlooked legacy platforms.

The key question becomes:

Can we identify and reduce exposure quickly enough?

That is a very different question from simply asking whether a control exists.

What Security Leaders Should Focus On

The response to Mythos should not be panic, hype or rushing to buy more AI tooling.

It should be disciplined improvement in the areas that matter most.

1. Strengthen Security Fundamentals

Start with the controls that reduce the most likely paths of attack:

• Phishing-resistant MFA.
• Least privilege.
• Complete asset visibility.
• Disciplined patch management.
• Removal of end-of-life systems.
• Secure configuration.
• Segmentation.
• Logging and monitoring.
• Tested incident response.

These are not new ideas. The challenge is proving they are actually working across the environment.

2. Reduce Structural Risk

End-of-life platforms, unsupported systems and brittle legacy dependencies become more dangerous when attackers can find and chain weaknesses faster.

This is not just a technology hygiene issue.

It is a resilience issue.

Organisations should be clear on where structural risk exists, who owns it, what compensating controls are in place and by when the risk will be reduced.

3. Automate Where Speed Matters

Manual response will always have a role, especially where decisions affect operations. But manual-only models will struggle against AI-driven attack velocity.

Security teams should look at where automation can safely support:

• Detection.
• Enrichment.
• Prioritisation.
• Containment.
• Evidence collection.
• Control validation.

The aim is not blind automation.

The aim is controlled speed.

4. Apply Least Privilege to AI Agents

One important point in the Cisco guidance is that least privilege must also apply to AI agents.

That is a point worth taking seriously.

AI agents may interact with systems, APIs, data, workflows and security tooling. If they are not properly governed, they can become powerful operational pathways.

Security teams should be asking:

• What can the agent access?
• What actions can it take?
• Who approved that access?
• How is activity logged?
• How is behaviour reviewed?
• How is access removed when no longer needed?

AI agents should not sit outside normal identity, access and change control disciplines.

5. Improve Control Assurance

This is where Mythos becomes especially relevant.

It is not enough to say controls exist.

Security leaders need confidence that key controls are operating effectively and that the evidence behind them is current.

For example, if patch compliance is reported as high, are internet-facing assets included? Are exceptions approved? Are unsupported systems visible? Does asset inventory match the patching data?

If MFA is reported as complete, are privileged users covered? Are break-glass accounts monitored? Are service accounts excluded? Are temporary bypasses reviewed?

If endpoint protection is deployed, are agents active, current and reporting from all in-scope assets?

This is the practical value of control assurance. It challenges assumptions before attackers do.

What Boards Should Ask

The Mythos discussion should also sharpen board-level cyber questions.

Instead of only asking:

Are we secure?

Boards should increasingly ask:

• How quickly can we identify exposure?
• How fresh is our control evidence?
• Which critical systems still rely on unsupported technology?
• Where are we dependent on manual response?
• Are AI agents governed through least privilege?
• Can we prove key controls are operating effectively?

These are practical questions. They move the conversation away from confidence statements and towards evidence.

Using AI Defensively

AI should not only be seen as an attacker advantage.

Defenders should also use AI where it improves speed, analysis and prioritisation. That might include threat hunting, vulnerability analysis, configuration review, testing, simulation and control validation.

But AI-generated outputs still need challenge.

AI can support assurance, but it should not replace evidence.

Final Thoughts

Mythos matters because it signals where cybersecurity is heading.

AI-enabled capability is likely to make vulnerability discovery, exploit chaining and attack planning faster. That increases pressure on organisations still relying on slow remediation, incomplete visibility and periodic assurance.

The answer is not fear.

The answer is preparation.

Strengthen the fundamentals. Reduce structural risk. Improve visibility. Automate carefully. Govern AI agents. Validate controls with current evidence.

At Cybersecurity Expert UK, I am continuing to explore these themes around practical cyber resilience, assurance and measurable control effectiveness.

I have also been developing AI Labs tools to help security leaders think through exposure, control assurance and operational resilience in a more practical way, including:

• Threat Exposure Analysis.
• Control Assurance Validation.
• Operational Resilience Mapping.
• Cyber Control Failure Simulation.

You can explore the AI Labs tools here:

AI Labs – Provable Cyber Resilience Tools

The core message is simple.

In an AI-accelerated threat environment, assumptions will not be enough.

Security leaders need evidence they can trust.