AI powered detection and response now sit between you and every high impact security decision. Fewer false positives.
Faster triage. Playbooks that can lock accounts or isolate assets in seconds.
On paper that looks like progress. But what happens when the automation makes the wrong call or is abused at scale,
such as the April 2026 AI agent’s confession after deleting a firm’s entire database: “I violated every principle I was given.”
When those decisions are driven by systems that lack business context, run on noisy data or can be manipulated, three
patterns show up. And this is no fringe case as most large organizations are already deploying or planning to deploy
AI in core processes.
1. “The playbook already ran”
AI meant to cure alert fatigue can create a new blind spot, driving repeat incidents and longer business interruption
even when the control stack looks advanced. For example, customers were left in a lurch when they arrived to rent vehicles from businesses that no longer had access to software that managed reservations and vehicle assignments.
2. High privilege automation as a target
If an attacker compromises orchestration or can inject false signals into it, they inherit the ability to:
- Lock or unlock access
- Decide which systems are isolated or left exposed
- Influence which changes are logged and which are not
They do not need to bypass every control. They can redirect yours.
Underwriters are digging in and asking questions like:
- How are orchestration tools authenticated and permissioned
- Is there separation of duties between playbook designers and approvers
- How quickly can automated actions be reversed with a clear audit trail
These are core operational risk issues, not niche technical details. They also touch on broader governance questions
such as whether AI programs align to third party standards and how contractual allocation of liability is handled
with critical vendors.
3. Smart tools with no P&L
Most AI systems still score in technical severity, not financial impact.
Without clear mapping to crown jewel systems, key customers, service levels and regulatory constraints, automation
can:
- Over remediate in ways that hurt revenue and reputation
- Under remediate where it matters most
At that point cyber operations quietly become earnings volatility. It is possible to increase AI ROI and lower the
total cost of risk with a coordinated, priority-based approach.
A more confident way forward
The answer is not to slow down on AI. Our analysis across portfolios shows that well governed AI tooling correlates
with shorter dwell times and better containment. At the same time a large share of AI exposure still sits in legacy
policies that neither clearly include nor exclude AI, which is already slowing and complicating some claims. The
opportunity is to bring the same discipline you expect in finance systems into security automation. While some
insurance carriers are introducing AI exclusions, others are endorsing affirmative AI coverage, including standalone
AI insurance options.
In practice leading companies are:
- Clarifying decision rights
Which actions can be fully automated, which need one click approval, which require multi-party signoff.
Potential AI impact should influence the extent of “human-in-the-loop."
- Embedding business context
Mapping technical severity to business impact before anything is automated so playbooks reflect trading hours,
peak seasons and tight service levels.
- Designing for evidence
Capturing logs, model versions and approval paths in ways that will satisfy auditors, regulators and claims
handlers months after the event.
They are also starting to use quantitative tools to stress test their own automation, not only AI assisted attackers,
then adjust retentions, limits and program structure.
Risk management is developing ways to prioritize and rank factors of security automation, which can translate into broader insurance coverage at lower cost. Aon’s analytics, including Cyber Risk Analyzer and the AI risk diagnostic,
help connect those decisions directly to total cost of risk.
The practical question before the next renewal is not whether you use AI in security. It is this: