Mandiant Outlines Guardrails for AI-Assisted Vulnerability Management

Article Details

Defense Focus

Safely integrate AI agents into vulnerability discovery and remediation using isolation, least privilege, deterministic validation, human review, and risk-based prioritization.

Detection Methods

• Normalize and deduplicate findings from EASM, CSPM, CTEM, and traditional scanners; enrich them with current asset context and threat intelligence to prioritize remediation.
• Use SAST and DAST for baseline code checks, and reserve intensive AI-agent audits for high-impact components where findings can be validated with observable outcomes.
• Require agents to produce reproducible test harnesses and run them automatically in isolated, monitored sandboxes with execution timeouts and iteration limits.
• Have a human review sandbox-confirmed findings for reachability, environmental context, and business impact.
• Monitor agent inputs, actions, reasoning, and outputs at runtime; use centralized logging to trace and restrict data flows to external endpoints.
• Monitor production workload behavior and the runtime libraries executing in production, not only static software inventories.

Data Sources

• Normalized vulnerability findings from traditional scanners, EASM, CSPM, and CTEM
• Updated CMDB and asset context, including exposure and data sensitivity
• Threat intelligence and exploitation telemetry, including EPSS probabilities
• Source code, dependencies, and SAST findings
• Sandbox execution results and deterministic test-harness outcomes
• Runtime workload and library telemetry
• AI-agent inputs, actions, outputs, and audit logs

Rule Types

• Deterministic SAST and DAST checks
• Deterministic policy-engine controls
• Schema validation for inbound application and API traffic
• Automated regression tests and exploit-validation test harnesses
• Weighted risk-based vulnerability scoring

Platforms

• Isolated, monitored sandboxes
• CI/CD pipelines and runners
• IDE-integrated development environments
• EDR platforms
• SOAR platforms

Defensive Actions

• Keep agents away from sensitive data; test with synthetic data in non-production environments and filter sensitive data and malicious prompt injections before they reach agents.
• Treat code, third-party dependencies, skills, and MCP integrations as untrusted inputs; vet tool integrity and monitor agent behavior at runtime.
• Run agents in isolated, unprivileged containers with dynamically limited privileges; use distinct, scoped machine identities and short-lived, repository- and branch-bound tokens.
• Use zero-data-retention agreements with LLM providers and establish authorized testing rules of engagement.
• Apply human-led threat modeling during design and architecture reviews, and red-team AI agents before deployment.
• Prioritize remediation using current asset context and threat intelligence; enforce patching SLOs and document risk acceptance for exceptions.
• Restrict remediation agents to proposing pull requests; run regression suites and the original proof-of-concept against proposed fixes, then require human approval.
• Use immutable audit logs, pin model API versions, and maintain deterministic regression suites and automated rollback capability.