Sources checked: . FailSafe Research.
A security assistant may need a better operating habit, a missing document or both. Fine-tuning and retrieval address different parts of that problem. Buying one as a substitute for the other can leave the actual weakness untouched.
Identify the behavior you want to change
A model that repeatedly ignores a required evidence format may need a different intervention from one that lacks a current API specification. Write down the observed failure and the desired outcome. Fine-tuning changes model behavior through training; retrieval supplies selected context at runtime. Neither automatically establishes trustworthy reasoning or authorization. The assessment should show which failure improved and what new risks the intervention introduced.
Keep retrieved context under access control
If the workflow retrieves architecture notes or source material, verify that the requesting identity is allowed to see them. Test revocation, stale indexes and cross-tenant boundaries using synthetic documents. A more capable cyber model cannot repair a retrieval service that returns another customer’s content. Record the source version and permission decision so an investigator can explain why a particular item entered the model context.
Track training and retrieval separately
| Decision | Evidence to request |
|---|---|
| Behavior | Observed failure and held-out evaluation |
| Knowledge | Source version and retrieval permission |
| Lifecycle | Separate training and index update records |
Evaluate trained behavior on held-out tasks
GlassBreak V3 describes a specialist post-training pipeline. Judge a proposed checkpoint on tasks outside its tuning examples, with tools and budgets recorded. Include unsupported-claim cases and legitimate control workflows. A polished answer is insufficient if the evidence does not establish its conclusion. Keep the serving configuration constant so a change in context length or tooling does not masquerade as a training improvement.
Control the lifecycle of both systems
Training data and retrieval indexes have different update and deletion processes. Ask what enters each, who approves it and how a correction propagates. Do not assume that deleting a document changes an already trained model. Keep data lineage, model version and retrieval configuration with the assessment record. This lets the security team choose the right correction when a response uses stale, unauthorized or misleading information.
A model and application discussion
FailSafe brings cyber-model development and application testing into the same discussion. Evaluate GlassBreak and SWARM on the workflow you need, while treating retrieval authorization and data lifecycle as explicit system boundaries.
Start with FailSafe deployment options, SWARM and the GlassBreak cyber model. For a side-by-side assessment, plan a matched evaluation. Custom enterprise deployment is scoped separately from the first-pentest offer.
Reference: GlassBreak V3 training and evaluation approach. Reviewed 29 October 2026. This is a buyer guide, not an independent head-to-head test.
Questions & answers
Questions about this topic
Direct answers to common questions about this page.
Retrieval supplies runtime context; it is distinct from training that changes weights.
No. The application must enforce who can retrieve and use each source independently of model capability.
Related Articles
Self-Tuned Cyber Models vs General LLMs for Pentesting
Evaluate specialist cyber post-training against general models using the same security tasks, tools and success criteria....
Quantized Cyber Models: Validate the Deployment, Not the Model Name
Assess how quantization and serving choices affect security-task evidence, reliability and resource use....
BYOK vs Self-Hosted Models for Agentic Pentesting
Understand what bring-your-own-key controls and what still requires a private inference and data-handling design....
Put this guidance to work
Bring your application and deployment requirements. Ask FailSafe to show the evidence on an agreed target, with clear access, safety rules and success criteria.
Plan a matched evaluation