IcePick · 01
The problem
LLMs and AI systems introduce failure modes that standard IT testing does not cover well: prompt injection, unsafe completions, policy bypasses, fragile guardrails, and gaps between what an organization thinks it controls and what the system actually does.
Scans alone are not enough. Neither are questionnaires. Teams need evidence-cited findings that connect technical results to organizational reality.
- Scans
- Static checks
- Questionnaires
Evidence-cited findings
IcePick · 02
What IcePick is
IcePick is Argus’s approach to LLM / AI-system stress-testing and reporting. It combines:
- automated / structured scans
- static checks
- organizational questionnaire inputs
…and turns gaps and hits into actionable, evidence-cited findings for your organization — suitable for security, risk, and engineering stakeholders who need to prioritize remediation.
IcePick is how Argus solves a recurring class of AI-security problems with rigor you can reuse — whether delivered as a focused engagement, a repeatable assessment motion, or both.
For practitioners: IcePick’s assessment motion is Garak-shaped.
IcePick · 03
What IcePick is not
- Not a guarantee that a system is “safe” or “unbreakable”
- Not an official compliance certification or legal stamp (EU AI Act, ISO, etc.)
- Not a substitute for your own risk acceptance and engineering ownership
- Not a black-box score without evidence
IcePick · 04
What you leave with
Clear findings tied to evidence, prioritized so your team can remediate and retest. Where organizational process gaps show up alongside technical hits, IcePick surfaces them instead of hiding them behind a single green/red badge.
IcePick · 05
Part of how Argus approaches AI security
IcePick sits naturally next to Argus’s adversarial ML and LLM-threat work. If you need a threat-response team, model-robustness work, or deeper red-teaming, we connect the dots — without forcing a false choice between “tool” and “engagement.”