One security leader put it plainly early in the research process. The biggest tension cyber leaders will face into 2027 is speed versus assurance on agentic AI adoption. Not because leaders disagree on whether to adopt it. Most already have. The problem is that the model the industry uses to govern it hasn't caught up.
For most of the last decade, AI security has been treated as a point-in-time exercise. Assess it. Sign it off. Move on. That model works when the thing being assessed is static.
Agents aren't static. They pick up new tools, new data sources and new instructions on a rolling basis. Their risk profile changes throughout their operational life, not before it. Several conversations converged on the same underlying problem. The industry doesn't yet have a shared answer for what continuous assurance actually looks like in practice. Who owns it. What it costs to run properly.
This isn't a hypothetical concern. It's already showing up in patch volumes. Leaders described a marked jump in vulnerability disclosures since major frontier model releases earlier this year. Some teams are now managing daily patch counts that would have been a monthly total eighteen months ago. The tooling that helps defenders find vulnerabilities faster is the same tooling helping attackers do the same thing.
If there’s one place this plays out most visibly, it’s agent identity. The market is evolving quickly, with vendors approaching the challenge from different points in the identity and security stack. Some focus on the user level. Some at the gateway level. Others are extending existing identity approaches to AI agents.
What came through clearly in these conversations is that CISOs recognise there isn’t yet a single, universally accepted approach. Several described being presented with different perspectives on the same challenge within a matter of weeks. Rather than creating more certainty, it highlighted just how many dimensions there are to consider. Vendors can provide valuable insight into the problems their solutions are designed to address, but the broader question is how those approaches fit together across the organisation.
The response, consistently, was to complement those conversations with peer exchange. Leaders called peers at other organisations and across industries to compare notes, understand how others were approaching the challenge, and sense-check where they had landed. Not because peer conversations replace vendor expertise, but because they provide another perspective on how the pieces are coming together in practice.
The most consistent theme across the research surfaced independently in very different settings. Long-established enterprise environments. Fast-scaling technology companies. Both described the same shift in security posture, from control to enablement.
One leader offered a good illustration. A departing employee sent a large volume of data home in the weeks before resigning. The system flagged it as unusual. The leader followed up directly instead of assuming the worst. The answer turned out to be mundane: five years of payslips, pulled together for tax purposes before the employee left. Nothing was blocked. Nothing was punished. The behaviour was visible enough to ask the question and legitimate enough to close it in one conversation. The same leader was candid that blocking that channel outright wouldn't have stopped the scenario it's meant to catch. It would only have removed the visibility that let them tell a legitimate case from a genuine one.
In fast-growing technology companies, the same principle showed up framed around commercial outcomes rather than data loss. Security's job isn't to slow down engineering velocity. That velocity is itself a competitive advantage. In both cases the underlying belief was the same. A security function built primarily around saying no eventually gets designed around. One built around visibility and fast, proportionate response earns enough trust to actually shape behaviour.
This isn't just a sentiment from research conversations. ASD's most recent Annual Cyber Threat Report recorded an 11% year-on-year rise in reported incidents and a 111% jump in critical infrastructure entities notified of malicious activity. It explicitly flagged AI as a factor enabling attacks at greater scale and speed. Its guidance to organisations is telling. Assume compromise. Focus protection on the assets that matter most.
That's close to what several leaders described doing already. Narrow the aperture to the systems and data that would genuinely hurt the business if compromised. Apply heavier controls there. Accept a lighter touch everywhere else. Not because everything else is unimportant. Because treating every asset as equally critical is no longer operationally realistic at the pace AI adoption is moving.
Taken together, these conversations point to something more specific than "AI is a growing risk." They point to an operating model taking shape in real time. Continuous rather than point-in-time assurance. Guardrails rather than gates. Trust built through peer verification rather than vendor assurance, because no single vendor currently has the full answer.
It isn't a finished model. Several leaders were candid that they haven't got the balance right yet. But the direction is consistent enough, across different industries and different scales of organisation, that it's no longer a fringe position. It's becoming the default way security leaders are choosing to govern something that refuses to sit still long enough to be governed the old way.
Photo by Fabio Sasso on Unsplash
For speaking and interview opportunities on cyber leadership topics, feel free to reach out to Maddie Abe.
Join us to share your insights at our upcoming events: