More results...
Cybersecurity has become a key priority for small and medium-sized businesses due to growing threats and wider AI adoption. An IDC survey of 2,200 SMBs in eight markets examined how organizations manage cyber risks, prepare for AI-related threats, and …
In this Help Net Security interview, Nick Nieuwenhuis, Cybersecurity Architect at Nedscaper, explains why cybersecurity has not delivered the resilience that decades of investment have promised. He argues that spending has leaned too heavily on technic…
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.
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:
The problem is not simply access. It is scale and autonomy.
That creates several immediate challenges:
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.
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:
That creates an entirely new form of runtime attack surface.
Recent research highlights risks including:
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.
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:
AI agents are fundamentally different.
They are probabilistic, adaptive, and capable of unexpected behaviour under changing context conditions.
This creates several assurance problems:
The challenge is not merely technical. It is operational.
Security teams increasingly face environments where AI functionality appears inside:
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 most significant AI security risk in many organisations is not the AI itself.
It is governance lag.
Technology deployment is moving faster than:
This creates a dangerous illusion of control.
Dashboards may still appear green while autonomous systems quietly accumulate:
Without strong governance, organisations risk repeating familiar mistakes:
The difference now is speed.
AI systems compress timelines dramatically.
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:
Most organisations cannot currently answer:
That must change quickly.
AI agents should be treated as managed operational assets with clear ownership and lifecycle governance.
Many AI deployments currently operate with excessive permissions for convenience.
That is unsustainable.
AI agents should operate with:
The principle of least privilege matters even more in autonomous environments.
The industry increasingly needs continuous validation rather than static approval models.
Security teams should focus on:
This aligns closely with broader Continuous Control Monitoring (CCM) approaches already emerging across cybersecurity assurance programmes.
Most governance structures were not designed for autonomous operational actors.
Boards, risk committees, and security leadership teams need clearer accountability models around:
The governance gap is becoming as important as the technical gap.
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:
That is ultimately the real challenge of AI security in 2026.
Not visibility alone.
But provable assurance.
Sources and further reading:
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.
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:
The problem is not simply access. It is scale and autonomy.
That creates several immediate challenges:
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.
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:
That creates an entirely new form of runtime attack surface.
Recent research highlights risks including:
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.
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:
AI agents are fundamentally different.
They are probabilistic, adaptive, and capable of unexpected behaviour under changing context conditions.
This creates several assurance problems:
The challenge is not merely technical. It is operational.
Security teams increasingly face environments where AI functionality appears inside:
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 most significant AI security risk in many organisations is not the AI itself.
It is governance lag.
Technology deployment is moving faster than:
This creates a dangerous illusion of control.
Dashboards may still appear green while autonomous systems quietly accumulate:
Without strong governance, organisations risk repeating familiar mistakes:
The difference now is speed.
AI systems compress timelines dramatically.
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:
Most organisations cannot currently answer:
That must change quickly.
AI agents should be treated as managed operational assets with clear ownership and lifecycle governance.
Many AI deployments currently operate with excessive permissions for convenience.
That is unsustainable.
AI agents should operate with:
The principle of least privilege matters even more in autonomous environments.
The industry increasingly needs continuous validation rather than static approval models.
Security teams should focus on:
This aligns closely with broader Continuous Control Monitoring (CCM) approaches already emerging across cybersecurity assurance programmes.
Most governance structures were not designed for autonomous operational actors.
Boards, risk committees, and security leadership teams need clearer accountability models around:
The governance gap is becoming as important as the technical gap.
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:
That is ultimately the real challenge of AI security in 2026.
Not visibility alone.
But provable assurance.
Sources and further reading:
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.
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.
Many organisations still manage cyber risk through processes designed for a slower environment.
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.
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.
Start with the controls that reduce the most likely paths of attack:
These are not new ideas. The challenge is proving they are actually working across the environment.
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.
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:
The aim is not blind automation.
The aim is controlled speed.
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:
AI agents should not sit outside normal identity, access and change control disciplines.
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.
The Mythos discussion should also sharpen board-level cyber questions.
Instead of only asking:
Are we secure?
Boards should increasingly ask:
These are practical questions. They move the conversation away from confidence statements and towards evidence.
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.
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:
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.
In this Help Net Security video, Jeffrey Wheatman, SVP and Cyber Strategist at Black Kite, discusses how organizations can identify and manage third-party cyber exposures before attackers exploit them. He argues that businesses should move beyond a dat…
Last week I joined fellow security leaders at CISO Inspire Summit North for a panel discussion on The Expanding Threat Surface: Adaptive Security Leadership for 2026 and Beyond.It was a timely discussion, because the challenge facing sec…
In this Help Net Security interview, Nuno Rodrigues Carvalho, Head of Sector for Incident and Vulnerability Services at ENISA, discusses the recent CVE funding scare and what it exposed about the fragility of global vulnerability disclosure infrastruct…
Security culture isn’t built by phishing simulations. In this Help Net Security video, Dan Potter, VP of Cyber Resilience at Immersive, argues that annual training videos and quarterly phishing tests happen in calm, controlled settings that tell …
One in five enterprise endpoints is operating outside a protected and enforceable state on any given day, according to device telemetry collected across tens of millions of corporate PCs. That figure, drawn from Absolute Security’s 2026 Resilienc…