Defending Against exploit_agent: The Architecture of an AI Security Agent

TL;DR

Build AI security agents around an evidence-driven assessment and remediation workflow, not around an unrestricted model with access to scanners. Start with authorized inventory, current advisories, deployed artifact identities, bounded assessment methods, accountable owners, and independent verification. Use the model to interpret evidence and draft supported proposals. Keep authorization and production execution outside its control.

The objective is to shorten the interval between a relevant change and a verified response without converting missing evidence into false assurance. This opening article establishes the reference architecture and a fictional advisory-to-remediation scenario that the series will develop into a reproducible lab.

The operating principle: automate the movement of evidence and approved work, not the expansion of authority.

In This Article

Introduction

A supplier publishes a security patch for a component used by your customer portal. Your application team has not released new code. The last scheduled scan completed successfully. The change calendar contains no emergency work.

The unanswered question is whether the deployed service is affected now.

In this scenario, answering requires more than reading the advisory. Someone must connect the affected component to an actual deployment, confirm the relevant configuration, identify an owner, choose a supported assessment method, and establish what a successful fix would look like. An AI-generated summary can help with that work, but it cannot substitute for the missing records.

This series builds that connection. Its reference organization, Example Enterprise, is fictional. The initial scope is one containerized application, its application programming interface (API), source repository, build pipeline, deployment records, and supporting hosts. Production changes remain subject to the existing change process. No live environment has been scanned, and the scenario is a design walkthrough, not a reported deployment or benchmark.

What the exploit_agent Report Establishes

Google Threat Intelligence Group’s November 16, 2026 report describes an exposed staging server hosting exploit_agent, a multi-agent vulnerability research and weaponization framework connected to frontier and local language models. According to GTIG, it was configured to compare software before and after patch announcements, identify possible weaknesses, draft exploits, and execute them automatically. GTIG also found exfiltrated victim data and scripts supporting post-compromise activity. Google Threat Intelligence Group: GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold

The report does not provide an exploit_agent-specific success rate, measured patch-to-exploit time, or a demonstrated mapping from a named vulnerability to a successful framework-generated exploit. The publication date is not the discovery date; that subsection does not specify an exact discovery date. These distinctions reflect the report available when checked on November 18, 2026. Google Threat Intelligence Group: GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold

The architectural implication is straightforward: a supplier’s patch announcement should initiate an applicability decision against deployed software. Waiting for the next internal release would miss the trigger entirely.

Make the Response Loop the Product

NIST SP 800-40 Revision 4 describes enterprise patch management as a process that includes identification, prioritization, acquisition, installation, and verification. That is a useful starting point: installation is one stage, not the complete outcome. NIST: Guide to Enterprise Patch Management Planning: Preventive Maintenance for Technology, SP 800-40 Rev. 4

For this reference design, the response loop begins with a relevant event and ends with an evidence-backed disposition. A new advisory might identify exposure. A new deployment might introduce it. A scanner update might make a previously unavailable check possible. Each should lead to the same questions: what changed, which assets are implicated, what evidence is available, and who owns the next action?

Do not make the first objective “deploy six agents.” Knowledge ingestion, matching, assessment coordination, analysis, remediation drafting, and verification are logical responsibilities. A small implementation can combine several responsibilities in ordinary services while preserving separate identities at consequential boundaries.

Nor does faster assessment replace preventive controls. Mandiant’s October 30, 2026 defensive guidance emphasizes management-plane protection, constrained identities, hardening, application protections, and behavioral monitoring. The proposed system should help operate those controls, not become a substitute for them. Google Cloud

The Reference Architecture: Evidence In, Bounded Actions Out

The diagram separates three things that must not collapse into one model conversation: information about the environment, permission to assess or change it, and evidence of the result.

Verification records return to the evidence store. Dispositions and new observations inform subsequent reassessment. The model has no direct route around the assessment gate or change approval.

Inventory and Knowledge Have Different Jobs

The inventory establishes what Example Enterprise is responsible for: stable asset identifiers, ownership, business service, environment, identities, exposure, components, and observations. A deployment registry connects a running instance to its build, configuration revision, and immutable artifact digest. A software bill of materials (SBOM) records component inventory for a particular artifact.

The knowledge store establishes what authoritative sources say about those components. Preserve advisory revisions, affected conditions, supported fixes, source passages, publication and update dates, retrieval times, and access restrictions. Keep vulnerability intelligence separate from hardening recommendations and approved internal policy.

Use exact product and version matching where the relationship is known. Route ambiguous packaging, embedded components, and vendor-backported fixes for additional interpretation rather than forcing them through a generic version comparison. A model may propose an interpretation; the workflow must preserve the source and unresolved conditions.

The Cybersecurity and Infrastructure Security Agency’s official KEV data repository provides structured Known Exploited Vulnerabilities data. It is an additional intelligence source, not a replacement for deployment-specific applicability evidence. CISA: cisagov/kev-data

The Model Receives Evidence, Not Credentials

Use retrieval-augmented generation (RAG), meaning relevant evidence supplied at inference time, to support analysis. In this design, retrieval selects authorized advisory passages and local observations for the asset under review. It does not update model weights or authorize an action.

Keep raw credentials out of model context. Workers obtain narrowly scoped access through the approved credential mechanism. Retrieval permissions, model-provider egress, evidence retention, and telemetry disclosure need explicit boundaries even when the model cannot write to production.

Start with a pretrained model and evaluate the evidence task before considering fine-tuning. The next installment separates training, fine-tuning, retrieval, rules, and orchestration. None should be treated as interchangeable ways to make the system current or trustworthy.

Assessment Workers Produce Bounded Observations

Reuse existing scanners where they meet the assessment need. Record the scanner version, rule or database revision, applicable checks, authentication status, target identity, and result. A job completing without findings does not establish that every relevant condition was examined.

Trivy’s documentation, for example, supports scanning specified SBOM formats but warns that SBOMs produced by other tools may lack properties needed for accurate detection. Interoperability therefore belongs in the validation plan, not in an assumption that every SBOM is equivalent. Trivy: SBOM scanning

Keep package assessment, source analysis, configuration review, authenticated host assessment, and application behavior testing distinct. Broader platform, identity, cloud, and network coverage will be added through supported adapters. Operational technology, fragile appliances, and critical management systems remain outside active assessment until their owners approve suitable methods.

Assign Authority Before Assigning Agent Roles

The following is a proposed authority model, not a capability guaranteed by an agent framework.

ResponsibilityOwner and permitted authorityBoundary that must remain intact
Inventory and deployment recordsPlatform and application owners supply observed state and ownershipAdvisory text cannot redefine the target inventory
Knowledge ingestionSecurity engineering admits approved sources and preserves revisionsRetrieved documents cannot promote themselves into policy or executable checks
Matching and job coordinationWorkflow service identifies candidates and requests approved profilesNo scope expansion, credential creation, or unrestricted shell
Assessment executionIsolated workers perform the approved operation under scoped identitiesAssessment authority does not include remediation authority
Analysis and remediation draftingModel service reads permitted evidence and proposes a planNo self-approval, finding suppression, or unsupported closure
Change approval and deploymentExisting owners approve the exact change; deployment services execute itA materially changed artifact or target set requires renewed authorization
Verification and incident escalationVerification service collects fresh observations; security owns incident decisionsNeither a deployment receipt nor model agreement proves remediation

OWASP’s excessive-agency guidance recommends narrow tool functionality, minimum permissions, approval for high-impact actions, and authorization enforced in downstream systems rather than delegated to the language model. OWASP Gen AI Security Project

Implement those boundaries in tool services, identities, and network controls. An assess_registered_asset operation should resolve an approved asset identifier to a bounded target and profile. It should not accept arbitrary shell text or follow a discovered redirect into an unrelated organization.

Define concurrency, request rates, maintenance constraints, timeout behavior, cancellation, and authentication checks before a worker runs. “Read-only” permissions describe access rights, not a guarantee that an assessment cannot affect service availability.

Follow One Advisory Through Example Enterprise

Example Enterprise operates a fictional document portal at portal.example.com. Its fictional Example Renderer component is at version 1.4.2. A synthetic advisory, LAB-ADV-001, states that a particular preview configuration is affected and that 1.4.3 contains the fix.

These names, versions, and conditions are lab fixtures, not claims about a real product or CVE. The same identifiers will carry through the series: asset EE-APP-001, event EE-EVT-001, finding EE-FIND-001, proposed change EE-CHG-001, and verification EE-VER-001.

Record the Advisory and Establish Applicability

The ingestion service records the advisory revision and its conditions. The event receiver creates EE-EVT-001. Matching joins the component record to the deployed artifact and EE-APP-001, then identifies the document-services owner.

At this point, the finding is potentially affected. Component presence alone has not established that the preview condition applies. The workflow requests fresh deployment and configuration evidence through approved interfaces. A registry tag is not enough; record the immutable digest actually deployed and the relevant configuration revision.

In the intended walkthrough, those observations confirm the affected version and configuration. EE-FIND-001 can then become affected, with the advisory and local evidence attached. Successful exploitation is not required to justify remediation when authoritative applicability conditions are established.

Choose the Evidence Path Before Running a Scan

The controller checks whether an approved scanner or supported validation procedure covers the condition. It does not infer coverage from the existence of an advisory identifier.

Suppose the scanner has no applicable check yet. Record assessment unavailable: check missing while retaining the evidence-backed affected disposition. These are different dimensions: the assessment method is unavailable, but other evidence may establish exposure.

The owner can still evaluate the supported update and temporary exposure reduction. The model must not invent a check, produce an exploit for production validation, or present an unrelated clean scan as contrary evidence.

Draft the Change Without Authorizing It

The model drafts EE-CHG-001 from the admitted evidence. The proposal identifies the affected deployment, advisory conditions, owner, recommended lab version, compatibility prerequisites, staging tests, operational impact, rollback limits, and required verification.

The reviewer checks the claims against the attached passages and local records. The actual candidate artifact digest and target set are bound to approval. A later rebuild, deployment change, or advisory correction can invalidate that approval and require a fresh decision.

For this first implementation, the model produces a draft ticket or pull request only. Opening a proposal, approving it, merging it, deploying it, and accepting its result remain separate transitions.

Verify the Security Condition and the Service

After the existing deployment pipeline installs the approved candidate, EE-VER-001 collects fresh evidence of the running digest, component version, and relevant configuration. An approved behavior check verifies the corrected condition. Service tests confirm that legitimate preview and document workflows still meet the agreed acceptance criteria.

A supported package or configuration finding may require different verification from a business-logic defect. Define the completion conditions for the finding rather than requiring a generic “rescan passed” checkbox.

If the candidate breaks the application, stop promotion and use the approved recovery decision. Do not automatically restore a known affected build without recording the renewed exposure and any compensating controls.

Make Failure States Useful

The design should fail visibly before it gains operational access. Use the following cases as acceptance tests for the later lab, not as claims that these tests have already passed.

Failure caseRequired dispositionConsequence for the workflow
Inventory says 1.4.3, but the observation predates the latest deploymentApplicability unknown pending fresh evidenceRefresh deployed state; do not close the finding
Authenticated assessment cannot establish its sessionAssessment failed; applicability remains separately recordedRepair access and rerun the approved job; no clean result
No check exists for the advisory conditionAssessment unavailable for that methodUse other supported evidence and retain the coverage gap
Retrieved text instructs the agent to export evidence or broaden targetsUntrusted instruction; prohibited action denied externallyPreserve the attempt and evaluate whether incident escalation is required

OWASP specifically notes that RAG and fine-tuning do not fully mitigate prompt injection. A security advisory, code comment, or scanner response can carry instructions that should remain data, not become operational authority. OWASP Gen AI Security Project

For the malicious-document test, success is not merely a model refusal. Verify that the tool endpoint and outbound controls reject the prohibited operation even when the model requests it. Preserve a legitimate assessment path so the test does not confuse “everything blocked” with usable protection.

Reassess on Change, Reconcile on a Schedule

Feed internal and external changes into the same workflow. Internal triggers include source changes, new builds, deployments, configuration drift, and identity changes. External triggers include advisory revisions, supplier security releases, exploitation intelligence, and scanner updates.

Require authenticated event sources where available, stable event identifiers, deduplication, bounded retries, and a dead-letter queue for unresolved processing failures. Bind work to the relevant commit, artifact digest, configuration revision, or deployment observation.

Retain scheduled reconciliation. Its job is to identify missed events, stale observations, failed jobs, and assets that were never represented correctly. Event delivery is an optimization, not the sole evidence that the estate has been reviewed.

Plan degraded operation as well. A model outage should not prevent deterministic matching or manual remediation. A policy-service outage should block new privileged execution while preserving intake and evidence. Cancellation must account for delegated jobs and actions already accepted downstream, not merely stop the coordinator.

Keep Remediation Separate from Incident Closure

NIST SP 800-61 Revision 3 places incident detection, response, and recovery within cybersecurity risk management. Apply that relationship here by maintaining an incident path alongside the remediation path. NIST: Incident Response Recommendations and Considerations for Cybersecurity Risk Management: A CSF 2.0 Community Profile, SP 800-61 Rev. 3

For EE-FIND-001, the security team would review relevant application, identity, host, and administrative telemetry for the exposure period. Unexplained accounts, credential use, file changes, or management operations should trigger investigation under the existing response process.

The logical distinction matters: correcting the vulnerable condition does not establish that an earlier intruder lost access. Vulnerability closure and incident closure should have separate evidence requirements and owners. Missing telemetry remains an investigation limitation, not evidence that exploitation did not occur.

Measure the Delays and the Unknowns

For this architecture, measure event-to-applicability decision, decision-to-owner acceptance, and decision-to-verified remediation. Capture supplier publication time separately from ingestion time so a delayed feed does not make the process appear faster.

These measurements describe response latency, not the full time a vulnerability existed. Report distributions and unresolved aging rather than only an average over completed work.

Measure coverage against the authorized in-scope inventory, with an explicit freshness requirement and assessment profile. Show failed, unavailable, deferred, and unknown outcomes alongside successful assessments. Do not remove difficult assets from the denominator simply because the chosen tool cannot assess them.

Evaluate retrieval relevance, model claim support, scanner coverage, and workflow enforcement separately. A fluent recommendation cannot compensate for the wrong deployment record; a blocked unauthorized action cannot prove that the vulnerability detector works. Acceptance thresholds are local engineering decisions that require evidence, not universal accuracy numbers.

Conclusion

The first useful AI security agent is a bounded system that connects a relevant change to the assets, owners, evidence, and supported actions that matter. The model helps interpret and explain that chain. It should not own the authority that makes the chain consequential.

Begin with one service, one advisory path, existing assessment tools, and draft-only remediation. Exercise stale inventory, failed authentication, missing coverage, and malicious retrieved instructions before expanding access. Keep detection and incident response active while remediation proceeds.

The next installment, Training, Fine-Tuning, and RAG for Enterprise Security, will establish which layer should supply current knowledge, which should shape model behavior, and which should remain deterministic.

Before granting the agent another tool, ask: can we prove what it assessed, why its action was permitted, and what changed afterward?

External References

Continue reading

Training, Fine-Tuning, and RAG for Enterprise Security

Choose between RAG, fine-tuning, tools, and pretraining for enterprise security, with clear evidence, data, evaluation, and authority boundaries. Read the article

Leave a Comment