AI-Assisted Vuln Management: A Blueprint for Safe Deployment

As highlighted in the Mandiant M-Trends 2026 report, the mean time-to-exploit (TTE) has dropped to -7 days—vulnerabilities are often exploited a week before a patch exists. To keep pace, security teams are integrating LLM agents into their pipelines for automated vulnerability discovery and remediation. But deploying privileged AI agents without mature integration processes introduces new risks. This guide from Mandiant Consulting provides actionable guardrails for safe AI-assisted vulnerability management.

Establish Operational Guardrails

Ground your approach in frameworks like the NIST AI RMF and OWASP Top 10 for LLMs. Google’s Secure AI Framework (SAIF) provides a practical path. Key controls include:

• Pre-agent data security: Enforce defense-in-depth with deterministic policy engines and guard models like Model Armor. Treat the codebase as untrusted input—indirect prompt injections can hide in comments.
• Zero data retention (ZDR): Ensure LLM providers don’t train on your proprietary code.
• Workload isolation: Run agents in unprivileged, sandboxed containers to contain blast radius.
• Red teaming: Subject agents to human-led red teaming before deployment.
• Least-privileged machine identities: Use short-lived, JIT tokens tied to human controllers.
• Supply chain resilience: Vet third-party MCP servers and orchestration frameworks for vulnerabilities.
• Toxic flow analysis: Monitor runtime data paths to prevent exfiltration.

The Need for Human-Led Threat Modeling

LLMs excel at syntax patterns but miss business intent. RAG helps but corporate docs are often stale. Human threat modeling using frameworks like PASTA remains critical to uncover architectural flaws.

Track 1: Enterprise Vulnerability Management

Address foundational security (secrets sprawl, missing MFA) alongside AI deployment. Use risk-based vulnerability management (RBVM) to prioritize: score = (W_vuln * S_vuln) + (W_asset * S_asset) + (W_threat * S_threat). Integrate EPSS for real-world threat context.

Track 2: Product Security & Development

LLMs are probabilistic—they struggle with complex logic split across chunks. They excel at binary oracles (e.g., memory corruption) but fail at architectural oracles (e.g., auth bypasses). Deploy selectively on:

• Memory-unsafe codebases (C, C++)
• Systems exposed to untrusted input
• Shared internal libraries
• Foundational security boundaries

Require agents to generate deterministic test harnesses. Only escalate to humans if the harness executes successfully. This reduces alert fatigue but doesn’t eliminate the need for skilled practitioners.

Remediation and Hardening

Use IDE-integrated agents for localized fixes with human review. For CI/CD runners, restrict to generating PRs that pass regression suites. Implement automated rollbacks and pin model versions to prevent drift. Maintain immutable audit logs for compliance.

Conclusion

LLMs accelerate vulnerability discovery but don’t replace secure-by-design principles. Phase out memory-unsafe languages and use LLMs to assist with code migration to Rust. The goal: build systems where entire vulnerability classes are eliminated by design.

Original announcement published on Google Cloud.