Five hundred people who do cloud security for a living were handed the same 23 problems and asked to sort them by how much risk each one actually creates and the real damage they are doing right now. That is the whole method behind the Cloud Security Alliance's Top Threats to Cloud Computing 2026, published on August 13, and eleven issues came out the other side.

The full 2026 ranking, with CSA's notes on 2024 placement:

1. Inadequate Identity and Access Management (Identity & Access Management in 2024 survey) (up from #2)

2. AI (Artificial Intelligence)-Enhanced Attacks (new)

3. Insecure Third-Party Resources (up from #5)

4. Insecure Interfaces and APIs (down from #3)

5. Misconfiguration & Inadequate Change Control (down from #1)

6. AI System Compromise (new)

7. Advanced Persistent Threats (up from #11)

8. Lacking Cloud Security Strategy & Governance (Inadequate Selection/Implementation of Cloud Security Strategy in 2024 survey/down from #4)

9. Insecure Software Development (down from #6)

10. Accidental Cloud Data Disclosure (down from #7)

11. System Vulnerabilities (down from #8)
‍

Two AI threats appear here for the very first time, one of them near the top. That part everybody expected. The more interesting part is what happened to everything that was already on the list, because almost all of it moved. Misconfiguration went from first to fifth. Advanced persistent threats climbed from eleventh to seventh.

Inadequate identity and access management took the top spot, and system vulnerabilities, the thing a large number of security programs are still measured on, came out last. AI did not just add two entries to this list. It reshuffled the whole thing.

‍

AI is not an emerging cloud security concern. It is already changing both how attacks are carried out and what organizations have to protect
Vic Hargrave, a lead author and chair of the Top Threats Working Group.
‍

‍
AI made everything vulnerable, so vulnerability stopped being the variable

Frontier AI models can now discover vulnerabilities across any target without human intervention. Anthropic’s Claude Mythos is the clearest example yet, demonstrating the ability to autonomously uncover zero-day flaws across software, operating systems, and complex codebases.

That basically means that everything is vulnerable, all of the time, and the gap between a flaw existing and somebody using it keeps shrinking toward nothing. Patching still matters and you should absolutely keep doing it. It just stopped being the thing that decides your outcome, because nobody gets to out-run a system that finds new doors faster than you can close them.

What the reshuffle landed on is access

Everything runs on access. Every human, workload, integration, third party, and now AI agent relies on open paths across the cloud to get the job done. Yet, like we mentioned here a few times before, over 96% of that access is never used, serving as wide-open paths for attackers who can now traverse every single path at machine speed.

So access is the attack surface of the AI era. The encouraging part is that unlike a zero-day, this one is yours. You granted it, it is written down in your own policies, and you can take it back.

This is why access sits at the very top of the list. An attacker’s foothold is only as dangerous as where it leads next. Looking at the CSA’s top threats, access is the single common thread beneath them all. Securing it pays down multiple risks at once, rather than tackling them one by one.
‍

Act secures the cloud from the ground up

That is the job we built Act to do, and it starts with seeing what access actually looks like in your cloud. We pull identity and network into a single Access Map, then read that map three ways: configured access, meaning what your policies allow on paper; effective access, meaning what is genuinely reachable once every permission and policy is evaluated together; and observed access, meaning what your logs have actually seen happen.

With all three in one place, least-privilege stops being a theory and turns into a guided action-plan you can work through. The access paths nothing depends on come out through your own native cloud controls, with nothing to deploy and no agents running in your environment, and the blast radius shrinks toward what the business genuinely uses. Adaptive controls keep access from sprawling, evolving right alongside your cloud. Teams continue shipping and granting permissions as needed, but all growth stays contained instead of drifting outward every quarter.

There is a nice asymmetry in all of this. An attacker, or an agent that has gone sideways, needs to find one path that works. You only have to remove the ones nobody is using. Of everything AI has reshuffled this year, that is the piece still firmly in your hands.