Why it matters
AI is compressing the time required to analyze large codebases. That matters most when defenders suddenly need to understand exposed proprietary source code or review a very large application before an adversary can turn code visibility into exploit paths.
The relevant capability is not simply an LLM reading files. Mandiant describes an orchestration system that structures threat modeling, discovery, analysis, validation and expert review into a repeatable pipeline.
The 100-plus result came from a specific incident-response case
Mandiant reported that its Agentic Vulnerability Discovery Harness, or AVDH, discovered more than 100 true-positive critical vulnerabilities in two days during a recent incident-response investigation involving stolen corporate repositories.
That is a notable operational result, but it should not be interpreted as a universal benchmark for every repository or model. Mandiant also reports using the harness across environments containing tens of millions of lines of code and thousands of analysis pipelines.
The harness is an orchestration system, not one autonomous prompt
AVDH chains specialized agents through a sequential workflow. The published architecture starts with threat modeling, then performs entry-point discovery and context enrichment before agents analyze access-control and data-flow hypotheses and send candidate findings through additional validation stages.
Environmental context matters. Mandiant specifically describes inputs such as asset inventories, software bills of materials, architecture documentation and threat intelligence as useful context for selecting relevant analysis rules and vulnerability patterns.
Human validation remains a deliberate quality gate
Confirmed machine findings do not automatically become final disclosures. Mandiant says human experts dynamically reproduce exploitation paths and execute proof-of-concept validation, discarding findings that do not survive testing or that encounter unseen compensating controls.
That distinction is important for enterprise adoption: agentic discovery can increase analysis throughput without removing the need for authorization, skeptical review, reproducibility and human accountability.
What changes for application-security teams
The operating model shifts from asking whether AI should participate in code review to deciding where it can safely multiply scarce expert capacity. High-value controls include limiting code access, maintaining trustworthy software and asset context, benchmarking the harness, preserving approval gates and measuring false positives and false negatives over time.
The strongest implementation is therefore a governed pipeline rather than unrestricted autonomous security testing.
The RFDELTA takeaway
Agentic application security is becoming an orchestration problem. The defensible advantage comes from combining model speed, deterministic workflow, organization-specific context and expert validation into an auditable system that can operate faster than manual review alone.
Watch the original Signal
The concise video version is designed for discovery; this page preserves the sourcing, caveats and deeper context.
Memorable path: https://rfdelta.com/028
Video transcript
AI just found more than one hundred critical vulnerabilities in two days. Mandiant says its Agentic Vulnerability Discovery Harness did it during an incident response investigation involving stolen corporate repositories. This was not one magic prompt. Specialized agents mapped entry points, traced data flow, challenged candidate findings, and pushed them through validation stages. Human security experts then reproduced exploit paths and discarded findings that did not survive testing. The shift is speed. Attackers can analyze leaked code at machine scale, so defenders need the same leverage without removing skeptical review. Agentic application security is becoming a production security pipeline. Follow RFDELTA for source led intelligence.
Frequently asked questions
What is AVDH?
AVDH is Mandiant's Agentic Vulnerability Discovery Harness, a structured multi-agent workflow for source-code vulnerability discovery and validation.
Did the AI autonomously prove every vulnerability?
No. Mandiant describes multiple machine-validation stages followed by human expert review and dynamic proof-of-concept testing before a finding is accepted.
Does the 100-plus result mean every codebase will produce similar findings?
No. The reported result came from a specific incident-response investigation involving stolen corporate repositories and should not be treated as a universal benchmark.
Primary sources
Continue exploring RFDELTA
RFDELTA Signals map the hidden systems, technology transitions and operational dependencies underneath fast-moving headlines.