top of page
Glossy 3D 48PEGASUS Logo.png

Cyber Risk from Early Hacking to Agentic AI

Defensive controls and resilience

A first-day incident decision card

FIELD NOTE Illustrative defensive workflow. Activate only within your authority and preserve safety and evidence.

Stabilize and establish command

Confirm operational impact, appoint an incident lead and identify affected service owners. Use contact details and a communications channel not dependent on the suspect environment. Record time-stamped facts, decisions and assumptions; designate who handles external communication, privacy and legal review. If safety-critical operations are involved, consult the responsible operations team before altering system state. [3]

Protect evidence appropriate to the situation and avoid uncontrolled access to sensitive records. Decide which access or service to limit, check continuity options, and record what the action is intended to accomplish. Not every event needs the same containment action.

Define a safe return

Before restoration, establish what clean means for identities, systems, data and dependencies. Restore limited functions first if safe, validate with accountable owners and monitor for recurrence. Communicate known facts, unknowns, expected service and the next update, and identify any reporting obligations with counsel. Avoid promises of total recovery before the evidence supports them. [13] [3]

Afterward, distinguish immediate fixes from systemic changes. Test whether the same failure would spread as far after changes. Give each corrective action a deadline, owner and an observable closure condition. A report without a retest is not a finished recovery process.

Identity must survive a primary outage

FIELD NOTE A breach of the central identity system can affect both production and response.

Prepare for the identity dependency

Map where staff, service accounts, vendors and agents authenticate. Flag which permissions can change records, credentials, payments or production settings. Require separate high-impact identities and minimize permanent privilege. Plan carefully tested emergency access that does not reuse the same failure path, with documented oversight and secure storage. [6] [5]

Confirm that key response, backup and communications tools can be reached independently under approved emergency conditions. Define who holds recovery materials and how access is reviewed after use. Remove stale and unnecessary access according to policy.

Verify recovery rather than assume it

Test revocation and restoration using authorized exercises. Record how long it took to restore validated access to each priority service. Check that emergency accounts cannot become unattended privileged entry points. During recovery, validate both direct logins and the relationships between services rather than only restoring one visible control.

For agent identities, require separate credentials, tight scope, short duration, and revocation through an authoritative service. Make sure people can halt downstream work even when the model interface is unavailable. Approval and identity changes should be reconstructible afterward through appropriate logs. [8] [3]

Prove that backups can restore services

FIELD NOTE An intact copy is not the same as a known-good and dependable service.

Define the sequence with operations

Agree which data, systems, identities, configuration, interfaces and communications are required for the first usable service. Set a service restoration target and a maximum acceptable data loss window. Document trusted versions and backup custody. Preserve additional boundaries between backup administration and normal production management. [13] [5]

A backup exercise should measure the time and staff effort to restore, plus application correctness. For a school or health system, meaningful metrics may include the ability to identify users safely, reconcile records and operate in an independent continuity mode.

Observe the difference between resumption and rebound

A service can launch before volumes return to normal. Measure reconciliation backlog, errors, user notification completion and support burden. Verify that software, identity and data were restored from sources supported by evidence. Preserve the ability to stop rollout if conditions change.

Test a partial restore and a wider restore with an important supplier withheld. After exercise, correct dependency maps and capacity assumptions. Close remediation only after someone not responsible for the original change verifies the evidence. The goal is to restore trust and service, not simply improve a downtime metric. [3] [13]

Design supplier controls around failure modes

FIELD NOTE Ask who can detect, revoke, restore and notify when the supplier is unavailable.

Obtain verifiable supplier evidence

Map dependencies by service, data and permission. Identify whether critical suppliers administer environments, process identities, deliver updates or hold recovery records. Ask how they isolate tenants, document authorized changes and support revocation or recovery. Ask what is actually included in audit and testing scope. [4]

Specify needed notification, response, evidence sharing, support and restoration terms before deployment. The party operating the infrastructure may not be the same party with legal reporting obligations or authority to suspend testing. Plan ownership explicitly.

Practice reduced dependency

When a single partner fails, practice returning to a limited but safe service. Confirm you can access essential data and records through approved paths; test replacement options and know expected limits. If a partner proposes an agent or new integration, reassess data boundaries and permissions instead of assuming it fits the existing agreement. [3] [8]

Track a portfolio of concentrated dependencies across the enterprise. Different suppliers may rely on the same platform, and separate products can share credentials or operators. Escalate cases where contractual remedies exist but no operational fallback can be demonstrated.

Gate AI actions at more than one layer

FIELD NOTE Scores and policies support release decisions; tested controls determine what can actually happen.

Three levels of authority

First, allow an agent to read a specified scope and make recommendations. Next, allow low-impact reversible changes within an explicit budget and log them. Finally, keep high-impact actions subject to fresh, independent human approval and a separate execution permission. Prefer denying unknown scope rather than inferring authority from user-facing language. [8] [6]

Approval needs context: show the proposed change, affected resources, source evidence, relevant policy and reversal plan. Don’t make the approver depend on a natural-language assertion for correctness. Verify the decision at the tool boundary and invalidate it when circumstances materially change.

Test failures and recovery

Use harmless fixtures and nonproduction records. Test appropriate stopping, appropriate continuation, stale approvals, tool unavailability, revocation, clock delays, and loss of logging. Use separate evaluators and affected-role reviewers when possible. Evaluate oversensitive refusals and missed stops separately; both affect business operation and safety. [9]

Require run budgets, scope-limited credentials, traceable actions, a human stop mechanism and operator visibility. Exercise the ability to terminate a running workflow without depending on its cooperation. Canary releases and controlled rollbacks should be appropriate to the operating environment; none removes the need for real ownership.

Run a nontechnical tabletop with useful measurements

FIELD NOTE Choose decisions that would be hard to make in the first hour, not a scripted hero outcome.

The scenario

A critical service is disrupted, user reports arrive through different channels, a vendor cannot confirm scope, and executives request a rapid return. Later, a purported internal instruction appears in an agent-managed workflow while the usual approver is unavailable. Keep the exercise focused on decision rights, facts and service continuity. No network access or attack simulation is required. [3]

Ask who can declare an incident, who manages continuity, how to confirm communications, when to suspend automated actions, and what evidence is required to restore each service. Record disagreements without treating them as errors before review.

Measures and follow-through

Track time to accountable command, time to confirmed alternatives, duration before high-impact automation was held, time to validated priority recovery, and extent of reliable records. Report medians only with adequate samples and show major tail cases. Describe whether a measure came from an exercise or actual incident. [13]

Take remedial actions with accountable owners, dates and retest requirements. Use variations in who is absent, which supplier fails and whether the usual communications system is trusted. This reveals whether backup authority and fallback modes exist beyond an idealized exercise.

Unknowns that can be monitored

FIELD NOTE Each uncertain claim needs evidence that would make you change the assessment.

Specify the uncertainty

For harmful actors using AI, distinguish provider-observed examples from incidence and outcomes. For more capable agents, look for replicated independent evidence conditioned on tool access, and for controls that keep working after model or workflow changes. For general intelligence, specify broad capability, transfer and reliability criteria before assigning a label. [8] [10] [9]

For quantum risk, monitor advances in error correction, logical operations, duration and validated application-scale resources. Monitor migration readiness separately: published standards, validated products, compatibility, inventory coverage and vendor deadlines. Progress in one column can change planning without proving another.

Public policy and institutional response

Track who has authority to report, decide, update standards and require remediation, distinguishing a proposal from an operative decision. Where law or standards differ by geography, ask appropriate specialists to confirm applicability. Avoid republishing sensitive incident facts or mislabeling a payment recovery as full operational recovery.

A change log should record the last verification date, denominator or scope, the decision that depends on it and the next scheduled check. Do not treat lack of public disclosure as proof that a capability or incident is absent. Use testable internal milestones even when the outside scenario remains uncertain.

Exhibit 14 · Evidence, inference and unknowns are different

Keep documented performance, conditional extrapolation and unresolved questions separate. Ask what evidence would change the judgment. Evidence from a narrow test does not establish broad autonomy, generalized intelligence, or the timing of a hypothetical discontinuity.

48PEGASUS Research conceptual exhibit. Not empirical data.

Conclusions and practical next steps

FIELD NOTE Build verifiable defenses that remain useful even if frontier forecasts are wrong.

What history and current data support

Cyber harm has expanded in some reporting measures, while the old trust assumptions still fail. Reported complaint losses do not quantify all attacks. Selected evidence of AI misuse supports preparations for speed and scale, not a claim that broadly autonomous attacks are ubiquitous. The important historical distinction is between limiting damage, bringing back a trusted service, and changing controls to prevent repetition. [1] [10] [3]

Start with essential services and identity dependencies. Test verified recovery and supplier fallback. Ask what stops an automated system from making changes when its user, tool, policy or evidence is unreliable.

What to prepare without pretending to predict

Inventory public-key uses and prioritize long-lived sensitive data and trust anchors. Plan interoperable migration against standards and validate performance and recovery. Track actual capability measures and emerging obligations without assuming a quantum breakthrough proves AGI or that AGI guarantees a singularity. [11] [12]

The strongest next investment is not a confident forecast. It is evidence that the organization can detect consequential changes, limit propagation, recover the services people need, and improve after the event. Assign owners, test the thresholds and report unknowns explicitly.

Previous chapter

Next chapter

Back to the report

Browse: White Papers · Reports · Downloads

bottom of page