Cyber Resilience In The Era Of Agents, by DOLIUM

For agencies running agentic AI, cyber resilience means restoring how their agents behave along with the data they hold. Backup, detection and identity recovery can bring systems back after an incident, yet none of them can confirm whether the agents switched back on will complete the work in the manner the agency intends.

The premise rests on four connected points that build on one another:

  • Agentic AI adoption has outpaced visibility, with research cited by Veeam showing that 88% of organisations use or pilot AI agents while more than 82% of enterprises have unknown agents in their environments.
  • An agent can only be said to have drifted if there is a readable standard it drifted from, yet most agencies hold that standard in policy documents, staff experience and customised Systems of Record.
  • DOLIUM keeps each organisation’s operating logic inside the enterprise, governs it and applies it to AI models at run time, giving leaders a reference point for what each agent is authorised to do, under whose authority and to what standard.
  • DOLIUM complements the cyber resilience stack rather than replacing it, protecting the operating logic that detection, backup and identity tools were never designed to hold.

The question recovery cannot answer

An agency can restore every system, dataset and identity after an incident while still having no reliable way of knowing whether the AI agents it switches back on will do the work the way it intends. That gap did not exist when systems ran only on code written and tested by people, yet it is now widening with every agent an agency deploys.

For many years, a successful recovery meant clean data, trusted identities and working applications. Once those three were back in place, an organisation could reasonably assume that work would resume as before. Agentic AI breaks that assumption in a way that most recovery plans have not yet accounted for.

Agents reason, plan and act across tools and data sources with limited human intervention, which means the way work is completed is no longer fixed in the application layer. Restoring the systems is therefore no longer the same thing as restoring the work.

Agentic AI and cyber resilience are now one conversation

Agentic AI and cyber resilience have become a single conversation because agents now read, change and act on enterprise data at machine speed. When Veeam launched Intelligent ResOps for Microsoft 365 on 15 October 2026, it cited its Data Trust & Resilience Report 2026, which found that 88% of organisations are using or piloting AI agents while only 7% qualify as truly AI-ready. The same release reported that more than 82% of enterprises have unknown AI agents operating in their IT environments.

The recovery data Veeam published earlier in the year is no more reassuring for leaders. Drawing on more than 900 senior IT, security and risk leaders, the report found that 90% of organisations were confident they could recover from a cyber incident, yet fewer than one in three ransomware victims fully recovered their data.

Australian agencies sit squarely inside this picture rather than watching it from a distance. ASD responded to 408 cyber security incidents reported by government entities in 2024–25, which was a third of all incidents it handled nationally.

The policy settings have moved accordingly. On 1 May 2026, ASD’s ACSC joined CISA, the NSA and cyber agencies from Canada, New Zealand and the United Kingdom to publish Careful Adoption of Agentic AI Services, which advises incremental deployment, continuous assessment against evolving threats and strong governance, accountability, monitoring and human oversight. The DTA’s updated Policy for the responsible use of AI in government now requires agencies to name an accountable owner for each in-scope AI use case, keep a use case register and assess each use case for risk, with the remaining requirements taking effect in December 2026.

What the resilience stack restores

The resilience vendors have responded to agentic AI with faster detection, immutable backup and identity recovery. Recent releases can find the agents running across an organisation’s cloud productivity environment, including agents nobody formally approved. They also show which sensitive data each agent can reach and let administrators reverse an agent’s changes file by file. Identity recovery now covers the directory services that both staff and agents rely on.

These capabilities matter because an agency that cannot detect a compromise, restore clean data or recover its identity infrastructure cannot recover at all.

What these tools restore, however, is the state of the systems. They establish what changed, who or what changed it and how to return to a known good point in time, but they cannot establish whether the agents now running on the restored systems are doing the work the way the organisation intends.

Drift needs a reference point

The question the resilience stack leaves open can be put simply: when an agent drifts, what exactly has it drifted from?

Drift is a comparison, which means it can only be measured against a defined standard. An agent that approves a payment, escalates a case or closes a request may be doing so exactly as the organisation intends, or it may be following a pattern introduced by a model update, a poisoned input or a configuration nobody recorded. Without a readable statement of how that work should be done, the two look identical in the logs.

The Five Eyes guidance points towards this gap when it urges organisations to plan on the basis that agentic systems may behave unexpectedly. Monitoring an agent’s actions is necessary, yet monitoring alone cannot say whether an action was correct, because correctness requires a reference point the organisation owns.

Recovering data and identities restores the systems, yet it does not tell an organisation whether the agents running on those systems are doing the work the way the organisation intends.

Where the standard for work actually lives

In most agencies, the standard for how work should be done is not held anywhere a machine can read. It is spread across three places that each capture only part of the picture:

  • Policy documents, which describe intent but rarely specify the decision rules, thresholds and exceptions that govern day-to-day work.
  • The experience of long-serving staff, who know which delegations apply, which cases need escalation and which rules bend under which conditions.
  • Customisations buried inside Systems of Record, where years of configuration have hardcoded business rules that few people can now explain.

Agents cannot be governed against a standard that cannot be read. When the standard lives in documents, memories and code, every agent deployment quietly encodes its own interpretation of how the work should be done, leaving the agency with no independent record to compare that interpretation against, either before an incident or after one.

The exposure compounds over time as the organisation changes around its agents. Staff rotate, systems are replaced and policy is rewritten, yet the agents keep running on whatever interpretation they were originally given.

DOLIUM holds the standard

DOLIUM holds that standard in a form both people and AI models can work from. Each organisation’s operating logic, meaning its decision rules, authority structures, processes and governance boundaries, resides within the enterprise, where it is defined, governed and auditable. That logic is applied to AI models at run time rather than embedded inside them.

This separation gives leaders a reference point for every agent: what it is authorised to do, under whose authority and to what standard. When an agent’s behaviour departs from that reference point, the departure becomes visible because there is something concrete to measure it against.

The same separation keeps the agency in control of its models. Because the logic was never embedded in a model, work can be rerouted to another model when one is compromised or withdrawn, without rebuilding the rules that govern the work. A compromised model therefore becomes a substitution decision rather than a re-engineering program.

For agencies working towards the DTA policy’s December 2026 requirements, the same structure means each AI use case can be traced to an accountable owner, an authority and a defined process.

Recovering behaviour as well as data

After an incident, recovery teams need to know which behaviour to restore as well as which data and identities to restore. DOLIUM gives the organisation a record of what was executed and against which process, which is the map recovery teams need to restore correct behaviour alongside the data.

That record changes recovery in three practical ways:

  1. Teams can identify which agent actions departed from the defined process during the incident window, rather than treating every agent action as suspect.
  2. Agents can be brought back against the governed version of the operating logic, so the restored environment runs to the standard the organisation set rather than to whatever state the agents had drifted into.
  3. Leaders can show auditors, ministers and the public what was executed, under whose authority and to what standard, which is the accountability that both the DTA policy and the Five Eyes guidance call for.

DOLIUM protects the operating logic

DOLIUM protects the asset those tools were never designed to hold, which is the organisation’s operating logic. The resilience stack restores the infrastructure while DOLIUM restores the way of working.

An agency that restores its systems without that logic has recovered its infrastructure and lost its way of working. As agentic AI moves from pilots into the core of government and private service delivery, the question for every agency is how it will use its own operating logic as the reference point for recovering, governing and scaling the agents it now depends on.

References

  1. Veeam, Veeam Intelligent ResOps for Microsoft 365 Helps Customers Safely Use AI, Cut Storage Costs, and Recover With Precision, 15 October 2026.
  2. StorageReview, Veeam Intelligent ResOps for Microsoft 365 Goes GA With File-by-File Undo for AI Agent Changes, October 2026.
  3. Business Wire, Veeam Research Finds AI’s Promise is Colliding with a Data and AI Trust Gap, 3 June 2026.
  4. Intelligent CIO, Veeam report reveals a market-wide shift from recovery confidence to proven data resilience, 15 April 2026.
  5. NSA, NSA joins the ASD’s ACSC and Others to Release Guidance on Agentic Artificial Intelligence Systems, 30 April 2026 (US time).
  6. A&O Shearman, CISA and ASD’s ACSC publish joint guidance on securing agentic AI systems, 20 May 2026.
  7. Cloud Security Alliance, Five Eyes Agentic AI Guidance: Translating the Careful Adoption of Agentic AI Services Joint Guidance into Actionable Security Programs, 4 May 2026.
  8. Digital Transformation Agency, Policy for the responsible use of AI in government, version 2.0.
  9. Digital Transformation Agency, AI policy update: strengthening responsible use across government, 12 January 2026.
  10. APS Academy, Policy for the responsible use of AI in government (v2.0), July 2026.
  11. ACS Information Age, Govt agencies fail to report cyber incidents, 2026.
  12. Australian Signals Directorate, Annual Cyber Threat Report 2024-2025, 14 October 2025.

This article was first published on DOLIUM’s LinkedIn page.

To explore how DOLIUM’s System of Work fits your operating model, book a briefing with the team.