Red-teaming for the era of LLMs and agents.
Prompt injection, jailbreaks, data exfiltration via tool calls, model denial-of-service, supply-chain attacks on training data. AI security needs a different toolkit — and we have built it for production deployments.
AI features have a new attack surface: prompts, retrieval inputs, tool calls, training data, model artefacts. We red-team your AI deployments — find the prompt injections that exfiltrate data, the tool-use chains that escalate privilege, the system-prompt leaks that betray architecture.
- ·AI threat model for your specific deployment (model, RAG corpus, tools, agents)
- ·Adversarial prompt library tested against your system
- ·Tool-call abuse + privilege-escalation testing (for agents)
- ·PII / system-prompt leakage assessment
- ·Training-data + supply-chain risk review
- ·Recommendations for guardrails, input/output filters, monitoring
- ·Re-test post-remediation
- ◇Custom adversarial prompt corpora
- ◇Garak (open-source LLM scanner)
- ◇NeMo Guardrails / Rebuff / Lakera (defensive)
- ◇Promptfoo (eval framework)
- ◇PyRIT (Microsoft) for systematic adversarial testing
Map your AI system's attack surface — inputs, outputs, tools, training corpus, model artefacts.
Run prompt injection, jailbreaks, indirect injection via retrieval, exfil via tool calls.
For agent systems: find chains that escalate privilege, abuse multi-step decisions, exfil via side channels.
Detailed findings + recommended guardrails + monitoring controls.
Validate fixes are effective and have not introduced regressions.
- ◆OWASP LLM Top 10
- ◆NIST AI RMF
- ◆MITRE ATLAS (adversarial AI)
- ◆EU AI Act high-risk categorisation
AI Threat Model + Findings Report — diagrams of your AI system's data flow + trust boundaries, a finding-by-finding writeup of every successful adversarial probe (with exact prompts that worked, conditions, impact), and a prioritized remediation roadmap.
Is this just normal pen-testing?+
No. Traditional pen-tests don't evaluate prompt injection, retrieval poisoning, model extraction, or agent-tool privilege escalation. This is a distinct discipline.
Do you do this for Claude / GPT / Gemini deployments equally?+
Yes. The attacks differ slightly by provider, but the categories — prompt injection, exfil, tool abuse — are universal.
Can you help us build the defenses?+
Yes — we cross-link with /ai/ai-development for the build side. AI red-team + AI dev together gives you a defensible posture.