One of the more interesting conversations coming out of RSA Conference 2026 had very little to do with autonomous AI, prompt injection, or whether frontier models are becoming dangerous.
The more meaningful focus for many security leaders was security and governance maturity, and whether most organisations are prepared for the level of visibility, access, and interconnectedness AI introduces into enterprise environments.
That was certainly the direction our recent webinar took with Rapid7. As a long-standing partner of The Missing Link, Rapid7 brings visibility across thousands of environments globally through its work in exposure management, threat detection, and vulnerability risk management.
Anthropic's Mythos research and Project Glasswing helped shape much of that thinking, raising important questions about how AI may influence vulnerability discovery, offensive security, and enterprise risk.
Beneath much of the discussion around AI capability was a far more immediate concern. While AI is introducing new attack surfaces and security considerations of its own, many of the most significant risks organisations face today are not entirely new. They stem from weaknesses in governance, identity, and architecture that become far easier to expose once AI systems are connected to enterprise environments.
As an industry, we've spent years talking about reducing attack surface, improving visibility, implementing Zero Trust principles, and tightening governance around sensitive systems and data. Many organisations are now discovering that introducing AI into environments that still lack those controls materially changes their exposure profile.
This is not because the model itself is inherently dangerous, but because the surrounding environment was never governed particularly well to begin with.
AI is exposing governance failures that already existed
When AI has more access than your employees
Across customer environments, The Missing Link's offensive security and security consulting teams are increasingly seeing AI systems granted access to repositories, sensitive documentation, cloud configuration data, and enterprise platforms that would traditionally be segmented or governed more tightly.
In some cases, those systems returned sensitive information, including data from internal records, without effective validation of whether the requesting user should have had access to it in the first place.
Architecture problems disguised as AI problems
These are not failures of the model but failures in architecture, access design, and governance discipline.
That doesn't mean AI-specific risks should be dismissed. Prompt injection, system prompt leakage, insecure model integrations, and other emerging attack techniques represent legitimate concerns. As discussed during our webinar, these issues are already appearing in real-world implementations and are becoming an increasingly important part of application security strategy.
What many organisations are discovering, however, is that governance failures often pose a more immediate risk than the model itself.
For years, organisations have worked to reduce excessive privilege across users, applications, and infrastructure. In many cases, AI integrations are quietly reintroducing those same risks through convenience, speed of deployment, or pressure to demonstrate AI adoption quickly.
The issue is not necessarily the AI system itself, but whether the organisation understands:
-
-
What the model can access
-
How permissions are governed
-
Where sensitive data exists
-
How outputs are monitored
- What happens when controls fail
-
These are architecture and governance questions, not AI questions.
The challenge is that AI significantly increases the potential blast radius when those questions have not been answered properly.
AI is accelerating attackers, not replacing them
The economics of offensive security are changing
There has been considerable discussion around whether AI will replace offensive security teams or automate attack workflows end-to-end.
That's not what we're seeing.
What AI is doing is reducing the effort required for many stages of the offensive workflow. Researchers can analyse larger codebases, accelerate reconnaissance, test attack paths more efficiently, and identify viable exploit combinations faster than before.
None of these activities are new. What has changed is the speed, scale, and accessibility with which they can now be performed.
This matters because it increases pressure on remediation. The industry hasn't historically struggled with identifying vulnerabilities. The more persistent challenge has been translating visibility into action.
Why exposure management is replacing vulnerability management
As AI-assisted research accelerates the identification of exploitable weaknesses and viable attack chains, organisations have less time to assess risk and respond.
Increasingly, the challenge is not a single vulnerability but how multiple weaknesses can be connected across identities, cloud services, applications, and infrastructure to create a viable attack path. This is one reason exposure management is rapidly overtaking traditional vulnerability management conversations across enterprise environments.
The same capabilities that allow AI to accelerate offensive research are also being applied to threat detection, exposure analysis, and security operations. As a result, the organisations that succeed will be those that can identify, prioritise, and remediate meaningful exposure faster than attackers can exploit it.
Identity and access are moving back to the centre
One of the more consistent patterns across customer environments is the renewed importance of identity as a central control point.
Historically, identity projects were often difficult to prioritise because they were organisationally complex and rarely perceived as transformative from a business perspective.
Once organisations begin integrating AI into enterprise repositories, cloud workloads, and productivity platforms, identity governance becomes much more difficult to separate from broader discussions around enterprise risk and security posture.
Every AI interaction is an identity problem
Every interaction with an AI system effectively becomes a query against the organisation’s underlying permissions model. Where that model is poorly defined, over-permissioned, or inconsistently governed, AI exposes those weaknesses very quickly.
Visibility is becoming just as important as control. Organisations increasingly need to understand not only what AI systems can access, but also how decisions are being made and what data is influencing those outcomes.
The controls moving back into focus
This is driving increased focus on:
-
-
Identity and Access Management (IAM)
-
Data Security Posture Management (DSPM)
-
Zero Trust architecture
-
Access governance frameworks
-
Cloud and AI security posture
-
Monitoring and control of AI system usage
-
The organisations managing AI adoption most effectively aren't slowing innovation. Instead, they're investing in foundational controls so that AI systems operate within clearly defined and well-governed boundaries.
Governance discipline will matter more than adoption speed
Common characteristics of high-risk environments
One of the clearest lessons emerging from RSA Conference 2026 is that AI is not destabilising mature environments but exposing environments where governance was already inconsistent or incomplete.
The organisations experiencing the greatest friction around AI adoption often share similar characteristics:
-
-
Overprivileged systems
-
Weak identity controls
-
Limited visibility into data access
-
Fragmented ownership models across teams
-
Inconsistent cloud governance
-
Reactive remediation processes
- Immature application security practices
-
AI amplifies these weaknesses because it operates across systems, repositories, data, and workflows simultaneously.
What secure organisations are doing differently
The organisations adapting most effectively are strengthening identity governance, maturing exposure management practices, accelerating remediation processes, and increasing visibility across cloud, application, and data environments before AI adoption scales further.
From The Missing Link's perspective, this is increasingly what separates organisations that are confidently adopting AI from those struggling to govern it. The challenge is rarely a lack of technology. More often, it comes down to visibility, ownership, and the ability to act on risk before it becomes exposure.
That may not be the most prominent AI narrative coming out of RSA, but it is likely to be one of the most consequential for enterprise security teams over the next several years. The businesses that succeed will not necessarily be the organisations adopting AI the fastest. More likely, they'll be the organisations capable of governing it effectively once it becomes embedded across systems, workflows, and decision-making processes.
Finding vulnerabilities faster only creates value if organisations can remediate them faster. As AI accelerates discovery, governance, visibility, and operational execution become increasingly important.
While AI introduces new attack surfaces that security teams need to understand and manage, many of the most significant risks emerging today are still rooted in familiar weaknesses. AI is increasing the speed at which those weaknesses are discovered, connected, and exploited.
RSA Conference 2026 may have been dominated by conversations about frontier AI models, but one of the most important takeaways was far less futuristic. The organisations that manage AI successfully will be the ones that can strengthen governance, identity controls, visibility, and remediation at the same pace they adopt new technology.
We explored those themes in greater detail during our recent webinar with Rapid7 and discussed what they mean for security leaders today.
For a deeper dive into Anthropic's Mythos research, Project Glasswing, adversarial AI, exposure management, AI-assisted vulnerability discovery, and the implications for enterprise security teams, watch the full webinar:
