ISC2 Secure London 2026

Summary and key takeaways from ISC2’s London-based conference.

The text "ISC2 Secure London 19 November 2026" on a white background.The word "secure" has a green gradient colour from dark to light green.The clock tower on the houses of parliament is show

This was my third year attending ISC2 Secure London, and the first conference conference I attended solely as a delegate this year. Being a London based event meant an early morning train, but that’s OK - there was a light breakfast on arrival 👍.

📝
You can read my write up of ISC2 Secure London 2025 here.

Format wise there was just one set of talks, so no risk of getting lost (even for me!) while finding the right part of the venue. We were again hosted at Convene in Bishopsgate, although a different Convene venue to last year.

Opening Keynote: Governing AI Before It Hacks You

Monica Verma, CCSP, from Monica Talks Cyber

Assume chaos, build resilience.

Monica took us through how "AI" has changed over 70 years from a tool that predicts, to saying things (generative AI), to now doing things (agentic AI). While agentic AI can be dangerous, the doomsday scenario where AI destroys the world is less likely. Pushing a doomsday narrative is dangerous in itself though. I'm not sure where I sit on that argument, but ultimately so long as humans retain control (and don't put weapons systems on networks accessible by AIs) then the chance of an AI controlled nuclear strike is less likely.

One example given of poor governance was where a company performed a trial using AI chat bots and tooling. For the trial, the AI was given access to all the company data, and then a selection of staff where given access to the tool. As a result, those staff had access to data they wouldn't ordinarily be given. This demonstrates that companies run AI agents based on flawed assumptions - in this case that AI runs with the same privilege and access as the person using the tool.

We've known that shadow IT [1] for years, and shadow AI is a real risk to organisations. Employees are seeing AI tooling as powerful at helping them get their jobs done, so if an organisation doesn't support staff with AI then the organisation will probably find its data going elsewhere. Knowing how shadow AI is being used, and which models, is a real challenge.

The UK Cyber Security and Resilience Bill: Next Steps to Get Your Organisation Ready

Panel session

I don't particularly enjoy panel sessions, and this was one of the weaker sessions of the day for me. The moderator allowed one speaker to dominate (by invitation!) so that lessened the panel's effectiveness, as did one of the speakers being unavailable last minute.

Importantly, the Cyber Security and Resilience bill is due to come into effect in 2027 assuming it passes into legislation this year. It aims to provide consistency for regulators and companies when it comes to compliance, and uses NCSC's Cyber Assessment Framework (CAF).

Another key point to note is that the reporting deadlines change from 72 hours down to 24. Unfortunately the speaker didn't say which deadline that related to...

Why AI in Cybersecurity Still Needs a Human in the Loop

Sponsored by Huntress

Huntress' sponsored session sounded promising but sadly could have been done in five minutes, which was a shame. I did like the comment that AI tooling can cause bottleneck shifting. Just because one task has sped up (your previous bottleneck), unless the whole system / process speeds up then all you've done is moved where the bottleneck is - not necessarily faster overall.

From Framework to Reality: Why “Good” Security Fails in Real Organisations

Renata Vincoletto, CISSP, CISO at Civica

Having worked in the public sector on and off for years, Civica was a recognisable name. Renata shared how Civica approaches security as an organisation, and reminded us that good security fails due to the gap between what's written down and what's really done. You may know this as "work as done vs work as written".

On policy, Renata highlighted that real organisations change faster than policies (the same is true for law). Given the rate at which technology changes, this can lead to some very outdated policies.

Risk owners are important, and should be named people rather than just a shared mailbox. Personally, I'd rather have risk tied to a post, in case the post-holder leaves, (so tied to the SISO rather than "Bethany Tomlinson" [2]) as that way it's probably easier to track responsibility. Ultimately though, someone has to be responsible for the risk and ensure it's managed - on that we both agreed.

This talk prompted me to check how my client's service desk handles password resets and MFA re-enrolments. We were reminded that having a report showing MFA was enabled for all users would be considered worthless if the service desk will let an impostor enrol a new MFA mechanism [3]. I'm pleased to say that my client has a good approach to that problem - maybe I'll have to add that to their pentest scope for next year.

Reviving the Hope in Automated Response: How CPR Decides Whether an Agent Should Act

Inbal Argov, VP of Product Strategy at Mate Security (sponsored session)

If a malware scanning tool detects a file is malware, can you delete it?
It depends - you have to consider false positives.

To say Inbal was passionate about her subject matter would be an understatement, and that gave her session plenty of energy and held engagement. Inbal put together a framework for Confidence, Precision, and Reversibility (CPR) which is how Mate's products use AI agents to take appropriate actions in response to incidents.

The basis of the framework is that we trust humans more with a task if it's something we've seen them do well before. She gave the example of being trusted to feed a friend's dog but not water their plants. Applying the same logic to AI agents, Inbal has worked to ensure that given a previously observed scenario, and giving the agent certain pre-approved actions, the agent will make the right choice each time. Alternative Competing Hypothesis is also used to support this, by forcing the agent to consider what supports its theory and what refutes it.

Action options are built by humans and then made available to the agent, and importantly these actions have reversible steps created at the same time. As a result, if the agent decides to take an action based on context, but later has to (or a human has to) undo that, it's safe to do so. Examples given included taking a server with one user offline (reversible: put the server back online), and deleting a suspected malicious database record after taking a snapshot of the database (reversible: restore the snapshot / import just that row).

Secure by Design? No, Secure by Default. Shift Left? No, Shift Everywhere.

Paul Holland, CISSP, Cyber Strategy Manager at Royal Mail

All AI is doing is changing the speed, it's not changing what we have to do as security professionals.

When you consider how AI has impacted software development, with it possible to write whole features that would have taken weeks in less than a day, it's no surprise that manual assurance cannot reliably keep up with modern practices. Some software companies are performing multiple releases an hour, let alone a day.

Paul advocates for adding "security to the flow, rather than adding more gates" as a way to ensure that security doesn't slow down the organisation. He also reminded us to log security concerns as tickets in the development backlog, so they don't get forgotten.

I'd heard of security champions before, people across the organisation that have an interest and possibly more training when it comes to cyber security. Royal Mail has Security and Privacy Risk Champions, that they call sparcs. I can definitely see having privacy in the mix being a useful addition.

AI Attacks and the Collapse of Defender Time

Justin Kohler, Chief Product Officer at SpecterOps (sponsored session)

AI makes exploitation inevitable.

Justin spoke last year, and his talk was engaging, so it was good to see him back again. Despite being a sponsored talk, Justin didn't repeatedly tell us how his product solves everything, and indeed highlighted that other products were available 😄.

BloodHound is a tool by SpecterOps (open-source version available, but check the licence before use) that shows how an attacker can pivot from a user to having control of a target environment. This all gets graphed nicely, and I really need to give it a go in my labs. Justin spoke about how it's important to understand attack paths for our organisations, and to then take steps to add choke points to break those chains. He did remark that breaking one chain didn't mean that you were safe - there'll be another way round!

Echoing a point from a presentation in the morning, Justin pointed out that the user invoking an AI agent may not be the identity that acts. Or put more simply "the AI agent may have more privilege and access than Jess from accounts that's using it" (my example).

Individually reasonable permissions produce unreasonable outcomes [when considered in the aggregate].

We have to consider the environment as a whole, rather than individually. Attackers don't just consider a single user, they consider what that user can do. That's how they build their attack chain, taking a user that looks unprivileged to one that can destroy the organisation.

We also need to remember that given the availability of AI tools, complexity is no longer a security control. While I'd argue it never truly was, complexity can slow down a human. An AI tool, mapping the environment, won't care.

The Cybercrime Files: Cases, Clues and Lessons from the Frontline

James Pamment, Detective Sergeant from City of London Police

James explained to us that the modern threat landscape is increasingly automated and fast, and scales well. This leads to fast moving attacks with potentially many victims. Of the attacks last year 88% started by phishing, further demonstrating that unsophisticated, high volume, steps can be successful for attackers.

When handling incidents it's important that for every decision made the rationale is also recorded. This will be useful for future audits (or potentially court). In James' policing, if something wasn't written down it may be considered by a court to not have happened, especially given that trials can often be years after the event and nobody's memory is perfect.

James' key takeaways for us:

  • Are we prepared to isolate our systems? 
  • Can we operate without technology? 
  • Do we know who is responsible for making decisions, communicating internally and externally etc? 
  • Don't forget to use knowledge-based proofs to defend against voice cloning calls ("where did we go for lunch last week?")

Cyber Careers in the Age of AI: Turning Disruption Into Opportunity

Kevin Fielder, ISAAP, ISSMY, CISSP, CISO at NatWest Boxed and Mettle

This was a "fireside chat", and those aren't usually my favourite sessions. Kevin was engaging though, and gave a refreshing set of viewpoints on how he manages a team and the importance of giving opportunities to those that want them.

Will I attend again next year?

For me, this conference offers good value for money at a convenient location. I'd certainly look to attend again.


Banner image: Cropped and edited ISC2 Secure London social media banner.

[1] Shadow IT - where staff bring in their own software or IT services that aren't managed, approved, or known about by the corporate IT team.

[2] Picked a random name.

[3] Can't decide where to put the punctuation in that sentence - sorry!!