AI Security: Threats, Defenses, and Recovery

AI SECURITY · ENTERPRISE ARCHITECTURE · RECOVERY

Secure AI. Control authority. Recover trust.

Protect AI systems, understand AI-assisted attacks, and build security agents with clear limits, trusted evidence, and tested recovery paths.

Choose the security decision in front of you

Start with the problem you need to solve. Each pathway connects threat understanding to practical controls and accountable operations.

IDENTITY · DATA · TOOL ACCESS

Protect AI systems

Design enforceable boundaries around agents, retrieval, connected tools, and the software they depend on.

Explore protection guides ↓

THREATS · EVIDENCE · EXPOSURE

Understand AI-assisted attacks

Follow incident reporting and research with explicit boundaries between observed activity, controlled evaluations, and forecasts.

Explore threat analysis ↓

INVENTORY · ASSESSMENT · VALIDATION

Build AI for security

Connect trusted security knowledge, asset ownership, bounded scanning, and model evaluation before agents receive operational access.

Explore security-agent guides ↓

CONTAINMENT · TRUST · RETURN TO SERVICE

Contain and recover

Limit compromised authority, resolve uncertain actions, and establish what must be proven before automated writes resume.

Explore recovery guides ↓

Start with three guides

Build the sequence from evidence, to authority, to incident response.

01 · UNDERSTAND THE EVIDENCE

AI-assisted campaigns, unsanctioned agent actions, and confirmed compromises require different evidence: provider reports, evaluation logs, and forensic records.

What AI cyberattacks have actually done

Compare incident reporting and controlled research while preserving their evidence limits.

Read the guide →

02 · CONTROL AUTHORITY

AI proposals and human policy meet at an execution gate, followed by a scoped executor, target system, and independent verification.

Keep AI from authorizing itself

Separate agent proposals, current authorization, controlled execution, and independent evidence.

Read the guide →

03 · PLAN THE RESPONSE

AI incident response distinguishes a stopped agent from verified containment of tokens and jobs, with independent identity and tool controls, recovery approval, and preserved evidence.

Contain compromised AI agents

Contain affected identities, tools, data, memory, and downstream actions.

Read the guide →

FEATURED WHITE PAPER

AI Agents and the Integrity of Enterprise Recovery

Explore false approval claims, recovered agent state, current authorization, and the evidence needed before protected writes resume. The paper includes a dated forecast and a proposed recovery exercise.

THE AI SECURITY LIBRARY

Follow a focused reading path

Expand a collection and follow its guides in order. Evidence reviews, forecasts, and proposed exercises are identified where they serve different purposes.

IDENTITY · DATA · TOOL ACCESS

Protect AI systems

Design enforceable boundaries around agents, retrieval, connected tools, and the software they depend on.

Explore protection guides (6)
  1. A governed RAG boundary covers source documents, processing, answers, chunks, embeddings, caches, logs, and backups, preserving policy across each data form.
    Your RAG Pipeline Is a Data Boundary, Not Just a Search FeaturePreserve restrictions and lineage across ingestion, embeddings, storage, revocation, and recovery.
  2. A verified requester's customer scope and current permissions determine the context allowed into retrieval and generation, while denied content remains behind the enforcement boundary.
    Retrieval Permissions Must Follow the User, Not the Service AccountCarry user permissions through retrieval, generation, caches, and conversation history.
  3. Versioned approved sources undergo claim-support, applicability, and disclosure checks before a bounded answer; unsupported claims remain unreleased and source viewing requires current access.
    A Citation Is Not Proof: Building an Evidence and Disclosure Control LayerCheck source identity, claim support, applicability, and disclosure permission.
  4. AI proposals and human policy meet at an execution gate, followed by a scoped executor, target system, and independent verification.
    The Architecture That Keeps AI From Authorizing ItselfSeparate agent proposals, current authorization, controlled execution, and independent evidence.
  5. Introductory visual for How to Secure an MCP Server with OAuth 2.1, Scoped Access, and Audit Logging.
    How to Secure an MCP Server with OAuth 2.1, Scoped Access, and Audit LoggingConnect token validation, narrow scopes, and authorization records.
  6. Repository files, MCP packages, and dependencies can reach coding-agent hooks, CI/CD credentials, and production unless execution and release authority are independently controlled.
    AI Development Supply-Chain Attacks: MCP Servers, Hooks, and CI/CDProtect coding-agent hooks, MCP servers, build identities, and release boundaries.

THREATS · EVIDENCE · EXPOSURE

Understand AI-assisted attacks

Follow incident reporting and research with explicit boundaries between observed activity, controlled evaluations, and forecasts.

Explore threat analysis (5)
  1. AI-assisted campaigns, unsanctioned agent actions, and confirmed compromises require different evidence: provider reports, evaluation logs, and forensic records.
    What AI Cyberattacks Have Actually Done: Separating Evidence From AssumptionCompare incident reporting and controlled research while preserving their evidence limits.Evidence review
  2. Separate ARTEX, exploit_agent, and CAIRN investigations inform reviews of application identities, supplier management, website and development dependencies, and industrial remote access.
    AI Cyberattacks in 2026: ARTEX, exploit_agent, and Enterprise RiskConnect reported agentic threats to identity, supplier access, containment, and recovery.Threat analysis
  3. ARTEX intrusion orchestration, exploit_agent vulnerability research, and CAIRN goal-driven operations call for identity, exposure, detection, and containment controls. Documented capability alone does not establish compromise.
    ARTEX, exploit_agent, and CAIRN: How Agentic Attack Tools DifferCompare documented framework roles, observed intrusion activity, and defensive controls.Framework comparison
  4. Illustrative path from AI-assisted reconnaissance through an OT access boundary to industrial controllers, with independent evidence required for physical impact.
    AI Attacks on Industrial Controllers: What the Evidence Says About Power-Grid RiskExamine industrial advisories and remote exposure with clear limits on what they establish.Evidence review
  5. An AI summary reporting approval confers no grant; a current authoritative approval record permits a controlled access grant.
    The Next AI Cyberattack May Begin With a False ApprovalRead a dated forecast with explicit conditions for assessing a future breach.Forecast

INVENTORY · ASSESSMENT · VALIDATION

Build AI for security

Connect trusted security knowledge, asset ownership, bounded scanning, and model evaluation before agents receive operational access.

Explore security-agent guides (6)
  1. Controlled security assessment connects current advisories and deployed assets to exposure evidence, a policy boundary, reviewed change, and verified results.
    Defending Against exploit_agent: The Architecture of an AI Security AgentStart with trusted evidence, asset inventory, bounded assessments, and verified changes.
  2. Current evidence and runtime context reach a model alongside an optional adapter; draft findings pass external validation and owner review.
    Training, Fine-Tuning, and RAG for Enterprise SecurityChoose the right combination of retrieval, tuning, tools, and pretraining.
  3. Declared and observed estate records are reconciled into owned assets, deployments, assessment requirements, current evidence, and visible gaps.
    Build the Asset Inventory Your Security Agents NeedConnect services and components to owners, evidence age, and assessment coverage.
  4. Advisories, threat signals, and baseline guidance pass source admission before becoming versioned evidence, permitted context, and reviewable proposals.
    Keep Security Agents Current with Trusted Advisories and Best PracticesBuild a controlled, versioned evidence supply from trusted security sources.
  5. An assessment request passes identity, scope, profile, and execution gates before a bounded worker returns a result or requires reconciliation.
    Build Security Agents with Controlled Access to Scanning ToolsBound tool access with assessment profiles, current authorization, and durable jobs.
  6. Retrieval, model, scanner, and workflow evidence remain separate and pass mandatory gates before a bounded next step.
    Test Your Security Model Before Giving It Operational AccessEvaluate model answers, retrieval, scanner coverage, and workflow enforcement separately.

CONTAINMENT · TRUST · RETURN TO SERVICE

Contain and recover

Limit compromised authority, resolve uncertain actions, and establish what must be proven before automated writes resume.

Explore recovery guides (6)
  1. AI incident response distinguishes a stopped agent from verified containment of tokens and jobs, with independent identity and tool controls, recovery approval, and preserved evidence.
    Enterprise AI Incident Response: A Runbook for Containing Compromised AgentsContain affected identities, tools, data, memory, and downstream actions.
  2. A restored runtime undergoes isolated validation while suspect memory and approvals remain quarantined; current authority and independent evidence govern a bounded return to service.
    AI Agent Disaster Recovery: Restore Trust Before AuthorityRequalify controls and accepted behavior before returning agent permissions.
  3. A current recovery authority holds restored generation 1 records closed until evidence is reconciled and new bounded execution is admitted.
    Restoring State Must Not Restore Authority: Independent Recovery Admission for AI AgentsCheck recovery admission before recovered agents regain operational authority.
  4. Reconcile before retrying: a recovered agent combines target state, operation records, and audit evidence to verify execution, reassess a retry, or hold an unresolved outcome.
    Reconcile Before You Retry: Recovering Uncertain AI Agent ActionsResolve uncertain outcomes before retrying work that could duplicate an effect.
  5. Before and after an incident, agent workflows, queues, context, and previous decisions require a new current-authority check before work resumes.
    When the Attack Returns After RecoveryExamine restored queues, shared memory, and stale approvals as possible recovery failure paths.Architectural analysis
  6. A restored AI platform proposes writes; an independent recovery control plane authorizes or denies them and requires a verified destination receipt.
    How to Verify AI Agents Are Ready to Resume WritesUse a proposed recovery exercise to test approvals, ownership, duplicate prevention, and stop controls.Proposed exercise

Latest AI security articles

New analysis, architecture guides, and recovery practices from the AI Security collection.

Browse all AI Security articles →

Connect security to the wider operating model.

Follow Enterprise AI for governance, platforms, and agents. Use Operations & Resilience for lifecycle control, service recovery, and the dependencies that support both.