Every vulnerability disclosure begins the same way: a raw report lands in a security team's inbox, often dense with technical detail, inconsistent in structure, and light on the standardized language that downstream databases and customers actually need. Transforming that material into a polished, accurate Common Vulnerabilities and Exposures (CVE) advisory is painstaking work — and for most Product Security Incident Response Teams (PSIRTs), it remains stubbornly manual.
That friction is now being targeted directly by AI-assisted automation. A compelling engineering approach has emerged in which a generative AI agent, grounded in authoritative vulnerability classification data, can produce a complete, structured CVE advisory draft in a matter of minutes from nothing more than a raw vulnerability report. The implications for security operations teams managing high-volume disclosure pipelines are significant.
The Hidden Cost of Advisory Drafting
The CVE advisory format is deceptively simple. The canonical template reads:
[PROBLEMTYPE] in [COMPONENT] in [VENDOR] [PRODUCT] [VERSION] on [PLATFORMS] allows [ATTACKER] to [IMPACT] via [VECTOR]
Behind that clean structure, however, lies considerable expert judgment. The PROBLEMTYPE field must map to a valid Common Weakness Enumeration (CWE) entry — one of hundreds of catalogued software weakness categories maintained by MITRE. The VECTOR field must align with a corresponding Common Attack Pattern Enumeration and Classification (CAPEC) entry, describing the specific technique an attacker would use to exploit the weakness. Neither lookup is trivial when you're working from a dense technical report written by a researcher who may not use MITRE's precise taxonomy.
Multiply this effort across an active vulnerability disclosure program — one that spans multiple product lines, ingests reports from external researchers and internal audits, and feeds into national databases including NIST's National Vulnerability Database, ENISA's European Vulnerability Database, and JPCERT/CC — and the manual overhead becomes a meaningful operational constraint. Delays in disclosure drafting create downstream delays in patching guidance reaching customers.
Why Standard LLMs Fall Short — and How RAG Fixes It
The instinct to throw a large language model at this problem is understandable. LLMs are fluent in technical language, trained on vast corpora of security documentation, and capable of producing advisory-style prose. The problem is accuracy at the classification layer.
Without grounding in authoritative source data, LLMs reliably produce what practitioners call hallucinated classifications — CWE and CAPEC identifiers that look syntactically correct but do not exist in MITRE's catalogues. This is precisely the failure mode documented in OWASP's Top 10 for LLM Applications. For a security advisory that will propagate to national vulnerability databases and inform enterprise patch prioritization decisions, a fabricated CWE-XXXX reference is not a minor embarrassment — it is a material accuracy failure.
Retrieval-Augmented Generation (RAG) resolves this by anchoring the model's output to a live, searchable knowledge base. Rather than relying on memorized training data, the AI agent queries indexed source content at inference time, pulling verified CWE and CAPEC entries before generating its response. The result is classification output that is both fluent and factually grounded.
The Architecture: Crawl, Index, Generate
The implementation pattern that makes this work involves two distinct phases before any AI generation occurs.
Phase One: Building the Knowledge Base
The MITRE CWE and CAPEC websites serve as the authoritative source. An automated web crawler is configured to scrape both catalogues systematically — seeding from high-level taxonomy views, following links matching vulnerability definition patterns, and extracting structured fields from each entry: descriptions, mitigations, consequences, and observed exploitation examples. This content is ingested into dedicated search indices.
Critically, the crawl is scheduled to run continuously, not as a one-time import. As MITRE publishes new CWE and CAPEC entries or updates existing ones, the knowledge base refreshes automatically. The AI agent's grounding data stays current without requiring manual intervention from the security team.
Phase Two: The Generative Agent
With the knowledge base in place, a custom AI agent is configured to accept raw vulnerability reports as input. At query time, the agent performs a semantic search against the indexed CWE and CAPEC data, retrieves the most relevant entries, and uses those as grounded context when drafting the advisory. The output includes:
The agent does not invent classifications. Every CWE and CAPEC reference it produces must exist in the indexed knowledge base — a hard constraint enforced by the RAG architecture rather than by prompt engineering alone.
What This Means for Security Operations
The practical outcome is a dramatic compression of the disclosure drafting timeline. Work that previously required a security engineer to manually cross-reference MITRE catalogues, interpret ambiguous researcher language, and produce consistent advisory prose now yields a structured draft in minutes. Human review remains essential — and appropriate — but the cognitive load shifts from generation to validation.
This matters most during high-velocity disclosure periods, when multiple vulnerabilities require simultaneous advisory preparation ahead of a coordinated patch release. It also reduces the risk of classification inconsistency across advisories authored by different team members over time.
What You Should Do
- Audit your current PSIRT workflow for manual steps that involve structured data lookups — CWE, CAPEC, and CVSS scoring are prime candidates for RAG-assisted automation.
- Prioritize grounding over raw generation when deploying LLMs in security contexts. Ungrounded models will hallucinate authoritative-sounding but fictitious classification identifiers.
- Automate knowledge base refresh cycles so your RAG indices track MITRE catalogue updates in real time, not at quarterly manual import intervals.
- Maintain human review as a mandatory gate before any AI-drafted advisory is published. The agent handles generation; your security engineers handle accountability.
- Document your AI pipeline for regulatory and audit purposes. As AI-assisted disclosure becomes more common, demonstrating the accuracy controls in your pipeline will matter to customers and regulators alike.
The vulnerability disclosure process has long been one of the most labor-intensive elements of enterprise security operations. AI agents grounded in authoritative taxonomies represent a genuine force multiplier — not a replacement for expert judgment, but a tool that removes the drudgery from classification work and lets security engineers focus on the decisions that actually require human expertise.
