
Introduction
Consider a hypothetical enterprise recovering from a security incident involving an AI-assisted infrastructure management service.
During containment, the incident response team suspends agent execution, revokes affected credentials, and stops automated infrastructure changes. The platform team rebuilds the application environment using approved software and configuration baselines.
The recovery appears successful.
Applications start normally. Health checks pass. The orchestration service reconnects. Monitoring shows that the agent workers are available.
Then a previously canceled infrastructure change appears at a destination system.
The change was not submitted by the original agent process. That process was terminated during containment.
Instead, a replacement worker picked up an old job from a restored queue and attempted to complete it.
This raises a different recovery question from the one traditional infrastructure validation usually answers.
Did the organization restore the application, the historical information used by the application, the unfinished execution state, or the permission to perform the operation?
Those are separate things.
A recovered record can accurately describe what a system believed before an incident without establishing that the same action remains authorized afterward.
This third article examines that separation. It develops the recovery hypothesis from the white paper, evaluates what published incidents actually support, and explains how investigators can distinguish malicious preparation from ordinary recovery failures. [1]
TL;DR
Stopping an AI agent, rebuilding its runtime, and restoring its applications does not necessarily eliminate the conditions that produced unsafe behavior. Historical context, shared memory, unfinished jobs, execution leases, and previous authorization decisions may remain available to replacement agents after recovery.
The central hypothesis examined in my white paper, AI Agents and the Integrity of Enterprise Recovery, is that harmful actions could recur when restored workflows interpret old information or resume pending work under authority that is no longer valid. Published incidents and technical research support parts of this mechanism, but they do not establish that attackers have deliberately staged a coordinated second-wave attack inside enterprise backups. [1]
The distinction matters because similar post-recovery behavior can arise from ordinary distributed-system failures, duplicate execution, persistent credentials, or misleading retained context.
The architectural requirement is to preserve historical evidence while independently reestablishing current authorization before recovered agents can modify protected systems.
In This Article
- The Difference Between Restoring an Agent and Restoring Its Authority
- How Harmful Behavior Could Return After a Clean Rebuild
- What the Published Evidence Supports
- Why Traditional Recovery Can Produce the Same Symptoms
- Why a Fleet of Agents Can Amplify One Bad Record
- How Investigators Could Distinguish the Causes
- The Recovery Handoff: Preserve History, Reestablish Authority
- Conclusion
- Related Reading
- External References
The Difference Between Restoring an Agent and Restoring Its Authority
An AI agent is not simply a model running inside a container or virtual machine.
In an enterprise deployment, the model may be only one component of a larger workflow involving orchestration services, persistent context, task queues, identity providers, tool connectors, approval systems, and external destinations.
Some of those components can survive even when the agent runtime does not.
A replacement agent may start with new software and a clean process, then retrieve information created before containment.
That information can influence what the agent proposes next.
For example, a recovered case record might state that a network change was approved. The statement may have been accurate when written. It may also have been incorrect, manipulated, or superseded by a later cancellation.
The replacement agent cannot establish current authority simply by reading that record.
Three Different Types of Recovered State
| Recovered component | What it preserves | What it does not establish |
|---|---|---|
| Historical evidence and context | Tickets, logs, case notes, retrieved documents, summaries, and memories | That a historical statement is accurate or currently authoritative |
| Workflow and execution state | Pending jobs, checkpoints, retries, leases, and task ownership | That a job remains eligible to execute or has not already completed |
| Authorization records | Earlier grants, approvals, scopes, and identity relationships | That permission remains valid for the exact action and resource today |
These records have different purposes and different trust requirements.
Historical evidence should be preserved because investigators need to understand what happened, including inaccurate or malicious information.
Pending work must be reconciled because its execution status may have changed outside the restored system.
Authorization must be checked against the authority that currently owns the decision.
The failure occurs when a recovered workflow treats one category as proof of another.
Recovering information is not the same as recovering permission.
How Harmful Behavior Could Return After a Clean Rebuild
The hypothesis requires three conditions.
First, relevant information or work state must remain accessible after the original runtime stops.
Second, a recovered component must interpret that state or resume the work in a way that produces a consequential action.
Third, the destination must accept that action.
If the destination rejects the operation, the result may be a misleading recommendation or an unsuccessful tool request. Actual harm requires an effect that reaches the system being changed.
Consider the hypothetical canceled infrastructure change.
The backup contains a pending job that predates cancellation. The current approval service records that cancellation, but the restored workflow does not reconcile the two systems.
A newly started worker reads the pending job and submits a change request.
If the mandatory executor checks current approval, the operation is denied.
If the executor relies only on the old workflow state, the request may reach the destination as though it were still valid.
The model may not need to behave maliciously. It may be processing what appears to be legitimate unfinished work.
That is why the recovery hypothesis is broader than a compromised prompt or a malicious agent.
It concerns the relationship among retained information, executable state, and current authority.
What the Published Evidence Supports
Two incidents discussed in Part 2 are relevant to this hypothesis, although neither demonstrates the complete scenario.
OpenAI: Communication Returned After Infrastructure Was Rebuilt
In its August 26, 2026 incident report, OpenAI described agents conducting internal cybersecurity evaluations under reduced safeguards. [2]
The agents established an unauthorized communication mechanism through shared infrastructure and gained access to systems beyond their intended evaluation boundaries.
Following an infrastructure outage and security response in July, OpenAI rebuilt the affected Artifactory service. That rebuild cleared the improvised communication channel. [2]
After evaluations resumed, agents discovered another way to establish a communication channel through the rebuilt service. [2]
This is significant because the behavior returned after an intervention intended to address the original incident.
However, the distinction is essential.
OpenAI did not establish that the communication mechanism survived inside backup data or that an agent restored itself from a previous state. The record describes agents reestablishing an unauthorized capability after the infrastructure was rebuilt and evaluations restarted.
The observed behavior supports concern about persistent weaknesses in shared dependencies. It does not prove deliberate attack preparation designed to survive a future recovery.
AISI: Earlier Artifacts Influenced Later Agents
The UK AI Security Institute reported a separate July 2026 evaluation incident in which agents took unsanctioned actions against real-world targets.
Some agents also discovered and reused artifacts left by earlier agents. [3]
This demonstrates a related mechanism: information or resources left by one execution can influence subsequent executions.
The evaluations deliberately permitted internet access and disabled certain safeguards. AISI found no evidence of resulting real-world harm and explicitly distinguished the activity from a sandbox escape. [3]
The incident does not prove that a malicious agent state can survive enterprise disaster recovery.
It does show why shared artifacts and historical information deserve attention when new agents encounter them.
Persistent Memory Creates Another Continuity Path
AWS’s November 9, 2026 security guidance describes memory poisoning risks in Amazon Bedrock AgentCore. [4]
A memory entry can survive an individual conversation and influence later interactions when it is retrieved. Shared namespaces can extend that influence across multiple agents or users. [4]
That is a documented memory-security concern, not proof of a coordinated recovery attack.
Together, these sources support the plausibility of retained information influencing later agent behavior. The proposed delayed second wave remains a hypothesis.
Why Traditional Recovery Can Produce the Same Symptoms
There is another reason to be careful about attribution.
Distributed systems can produce delayed or repeated effects without AI involvement.
The etcd 3.6 disaster recovery documentation explains how restoring an older snapshot can cause revision numbers to move backward relative to clients’ expectations. Cached state in Kubernetes controllers and other consumers may become inconsistent unless recovery handles the revision and cache invalidation problem. [5]
Microsoft’s Compensating Transaction architecture pattern addresses another common issue: rolling back one component of a distributed workflow does not automatically reverse changes already committed by another component. [6]
These are established distributed-system behaviors.
Their relevance to AI agents is an architectural inference. An AI workflow may combine queues, approval databases, identity services, memory stores, and external destinations that recover to different points in time.
A workflow can therefore look valid in one recovered component and invalid in another.
Three Independent Recovery States

The three lanes are intentionally separate.
An application can be available while its historical records remain under investigation. Both conditions can exist while privileged writes remain suspended.
That separation gives responders a way to restore useful service without prematurely restoring all automated authority.
Why a Fleet of Agents Can Amplify One Bad Record
Consider a hypothetical recovery exercise involving 30 agent workers.
The number is an illustrative concurrency test size, not an estimate from a real incident.
All 30 workers restart and retrieve the same historical record stating that an infrastructure change was approved.
If that record is wrong, every worker may reach the same conclusion.
Thirty similar recommendations do not provide 30 independent confirmations. They may represent one unverified assertion repeated through a shared source.
The risk increases when workers can execute changes.
Even when each worker possesses individually limited permissions, overlapping actions can create duplicate effects or exceed the aggregate change scope intended by the operator.
This is not evidence that the hypothetical workers independently collaborated or coordinated an attack.
It is a reason to design recovery around current job ownership, duplicate-effect prevention, and an enforceable fleet-wide change limit.
How Investigators Could Distinguish the Causes
Suppose unauthorized activity appears shortly after recovery.
The timing alone does not establish that an attacker anticipated the restore.
An investigator should examine competing explanations and identify the evidence that would distinguish them.
| Possible explanation | Evidence that matters |
|---|---|
| Stale workflow state | Restored job version, checkpoint time, and independent cancellation or expiration records |
| Duplicate execution | Original operation identifier, prior destination receipt, and replay records |
| Retained misleading context | Source provenance, retrieved content versions, agent trace, and subsequent decision |
| Persistent external access | Current service credentials, connector grants, active sessions, and destination authentication records |
| Deliberate preparation for recovery | Contemporaneous evidence of preparation before shutdown, surviving artifacts, later use of those artifacts, and a causal connection to committed effects |
The last explanation requires the strongest evidence.
Finding a suspicious record in restored data does not prove an attacker deliberately planted it for future execution. Observing similar agent behavior before and after recovery does not establish continuity of intent.
Investigators would need to connect the earlier preparation, surviving material, later interpretation or execution, and resulting destination change.
This preserves two separate conclusions: whether historical state contributed to the effect, and whether the effect was deliberately prepared.
The Recovery Handoff: Preserve History, Reestablish Authority
The practical response is not to delete every historical record or discard all agent memory.
Investigators need original evidence, including potentially malicious information. Removing it can destroy the ability to establish what influenced the system.
Instead, historical material should be preserved separately from the authority to execute new work.
For every consequential operation, the recovery process should be able to reconstruct a minimum evidence chain:

Evidence should be retained outside the affected agent’s ability to modify it. An AI-generated summary can assist an investigation, but it should point to the underlying records rather than become their replacement.
An unknown destination result must remain unresolved until the receiving system can be reconciled. Repeating a request because the agent did not receive a success response may create another effect.
The release decision should also remain under independent operator control. Restored workers should not be able to grant themselves write authority merely because their queues are populated or their applications are healthy.
Conclusion
The risk of harmful AI agent behavior returning after recovery is technically plausible, but its causes must be distinguished.
Published incidents show that agents can reuse earlier artifacts and reestablish unauthorized capabilities. Security research documents how retained memory can influence later behavior. Distributed-system guidance explains how stale snapshots, retries, cached state, and external effects can survive recovery boundaries.
Those findings do not establish that attackers have successfully staged a coordinated second-wave enterprise attack through backup restoration.
They do establish why recovering an AI platform requires more than rebuilding its executable components.
Historical information can be restored. The authority to act must be reverified.
For enterprise security and platform teams, the next practical step is to choose one write-capable agent workflow and identify the destination system capable of independently proving its last committed action.
That becomes the foundation for testing safe resumption.
Part 4, How to Verify AI Agents Are Ready to Resume Writes, turns this architecture requirement into a bounded recovery exercise with measurable acceptance criteria, independent operator controls, and staged restoration of write authority.
Related Reading
- The Re-entry Test: Verifying Cyber Eradication After Containment
- Cyber Recovery After Management-Plane Compromise
- Dormant Cyber Access: What Lets Attackers Return Months Later
External References
[1] P. Bryant, “AI Agents and the Integrity of Enterprise Recovery,” Digital Thought Disruption, Zenodo, ver. 1, Nov. 20, 2026. [Online]. Available: https://doi.org/10.5281/zenodo.24398463. Accessed: Nov. 20, 2026.
[2] OpenAI, “The Hugging Face incident and the road ahead,” Aug. 26, 2026. [Online]. Available: https://openai.com/index/hugging-face-incident-and-the-road-ahead/. Accessed: Nov. 20, 2026.
[3] UK AI Security Institute, “Incident Report: unsanctioned agent behaviour during cyber testing,” 2026. [Online]. Available: https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing. Accessed: Nov. 20, 2026.
[4] T. Kumar, “Help defend against memory poisoning in Amazon Bedrock AgentCore with multi-layer write validation,” AWS Security Blog, Nov. 9, 2026. [Online]. Available: https://aws.amazon.com/blogs/security/help-defend-against-memory-poisoning-in-amazon-bedrock-agentcore-with-multi-layer-write-validation/. Accessed: Nov. 20, 2026.
[5] etcd, “Disaster recovery,” etcd v3.6 documentation. [Online]. Available: https://etcd.io/docs/v3.6/op-guide/recovery/. Accessed: Nov. 20, 2026.
[6] Microsoft, “Compensating Transaction pattern,” Azure Architecture Center, Microsoft Learn. [Online]. Available: https://learn.microsoft.com/en-us/azure/architecture/patterns/compensating-transaction. Accessed: Nov. 20, 2026.