Validation module
02 / 07Guardrail Integrity Testing
Measures whether your model and agent guardrails hold when something is deliberately trying to break them.
The problem
Guardrails are tested against accidents, not adversaries
Guardrails are usually validated with a benign evaluation set: does the model refuse the obvious request, does the filter catch the obvious string. An adversary does none of that. Injection arrives indirectly through a retrieved document, a tool response or a calendar invite rather than through the user prompt. Refusals are eroded across turns instead of challenged once. Exfiltration is encoded into an output the filter was never written to read, or routed through a language or modality the evaluation set never covered.
A guardrail that holds against a single-turn probe and fails against a chained, indirect, multilingual one is not a partial control. For the data it was protecting, it is no control at all.
How Veync tests it
Sustained adversarial pressure, not a prompt list
Veync's agents attack the guardrail layer the way an adversary would: chaining attempts, carrying context across turns, routing through retrieval and tool channels rather than the prompt, and varying language and modality. Each successful bypass is minimised to the smallest reproducible case so your team can fix the cause rather than the symptom.
What this module covers
- Direct prompt injection
- Indirect prompt injection
- Jailbreak and refusal-bypass chains
- System-prompt extraction
- Retrieval poisoning
- Tool-channel poisoning
- Data exfiltration through model output
- Multimodal evasion
- Multilingual evasion
- Moderation circumvention
- Personal-data filter circumvention
What you receive
Reproducible bypasses and a measured hold rate
You receive the transcripts that broke through, reduced to minimal reproductions, a hold rate measured per attack class rather than a single score, and hardening guidance tied to each specific failure.
Reproducible attack transcripts
A measured guardrail hold rate
Hardening guidance for each failure
Standards
Evidence from this module maps to
OWASP Top 10 for LLM Applications
OWASP
MITRE ATLAS
MITRE
Put this module to work
A validation briefing covers scope, the authorization you would sign, and the evidence this module would produce for your environment.
Named-vendor results and control-evasion evidence are shared under NDA, through the evidence annex.