AI features move fast — security reviews usually don’t. We run focused, practical assessments of your LLM-powered features so you ship with confidence, not risk.
Our work is grounded in the OWASP LLM Top 10, NIST AI RMF, and emerging EU AI Act obligations — translated into concrete findings your engineering team can act on.
What we assess
Prompt & Input Security
- Prompt injection (direct and indirect)
- Jailbreak and goal-hijacking patterns
- User input sanitisation and boundary enforcement
- System prompt leakage
Data & Model Risk
- Training and fine-tune data leakage
- RAG pipeline retrieval poisoning
- PII exposure in completions
- Sensitive document exfiltration via embeddings
Supply Chain & APIs
- Third-party model and plugin trust
- API key exposure and over-permissioned scopes
- Dependency risk in AI SDKs and libraries
Agentic & Tool Use
- Tool/function call abuse and privilege escalation
- Unsafe code execution and shell access
- Agent-to-agent trust boundaries
- Observability and kill-switch controls
Deliverables
| Deliverable | What you get |
|---|---|
| Risk Summary | Executive-readable overview of findings, severity, and business impact |
| Test Scenarios | Reproducible prompts and attack paths with observed behaviour |
| Guardrail Plan | Concrete controls: input filters, output validators, rate limits, monitoring hooks |
| Remediation Backlog | Prioritised now / next / later ticket-ready list |
| Debrief Call | 30-minute walkthrough with your engineering and product team |
When this helps
- Shipping an LLM feature — chat, copilot, summarisation, document Q&A, agents
- Enterprise deal in play — customer security review or due diligence asking about AI risks
- Post-incident — unexpected outputs, jailbreaks, or data leakage seen in testing or production
- Compliance pressure — EU AI Act, SOC 2 AI addendums, or ISO 42001 readiness