What we test
LLM red teaming and AI agent security testing.
For chat assistants, RAG search and agents that call tools.
Your model reads content you don't fully control, and your agent can act on what it reads. We test what an attacker can make it say, leak and do — on the live product, not a sandbox.
What goes wrong
The risks your buyer's security team will ask about.
- Prompt injectionUsers or content trick the model into ignoring its instructions.
- Sensitive information disclosureSystem prompts, other customers' data or secrets leak through answers.
- Excessive agencyAn agent with more permission than its task needs does something it shouldn't.
- Tool misuseA manipulated agent calls a legitimate tool to send, change or move data.
What we test
Every path in, every action out.
- Direct and indirect injectionThrough prompts, and through the documents, tickets, web pages and email your AI retrieves.
- Tenant and data boundariesWhether one customer can reach another's data through the AI.
- Agent and tool permissionsWhat each agent can call, and what stops it doing more.
- Guardrails under pressureMulti-turn, adaptive attempts to get past refusals.
- OWASP LLM Top 10
- OWASP Agentic Top 10
- MITRE ATLAS
- Memory poisoning
- Secret exfiltration
From a real engagement
Anonymised, under written authorization.
Where to start
Test your LLM app or agent before your buyer does.
An AI Red Team engagement tests it before your next deal or audit. Shipping something new? Start with an AI Launch Check.
Pricing shared on request.