AI & LLM Security
AI models are susceptible to new attacks like prompt injection. Securing them is vital for safe deployment.
Who this is for
Engineering and security teams preparing for a release, a customer security review, or a procurement questionnaire — and anyone who needs to know what an attacker could actually reach.
Methodology
What We Test
- Prompt injection and instruction override paths
- Training data exposure and sensitive data leakage
- Model output manipulation and response poisoning
- Plugin, tool, and agent integration abuse
- Authentication, authorization, and tenant isolation flaws
- Insecure model configuration and deployment weaknesses
How We Test
We start from attacker-controlled inputs, not trusted prompts. We actively bypass safety controls and alignment assumptions. We chain prompt abuse with application and API flaws. We validate real data access and execution impact. We escalate from model misuse to application or account compromise.
What You Receive
- Exploitable attack paths, not theoretical risks
- Clear reproduction steps with payload examples
- Impact assessment tied to data access or control
- Remediation guidance aligned to exploit paths
How an engagement runs
-
Scope & authorisation
We agree exactly what is in scope, what is explicitly out, and the conditions under which we stop. Nothing starts without written authorisation from someone able to give it. NDA first if you want one.
-
Access & prerequisites
We tell you up front what we need — test accounts at each privilege level, a non-production environment where that applies, documentation, and a named technical contact. Missing prerequisites are the usual reason assessments slip.
-
Reconnaissance & threat modelling
We map the real attack surface and decide which paths are worth the time, based on how your system is actually built rather than a generic checklist.
-
Manual testing & validated exploitation
Testing is hands-on. Findings are reproduced before they are written up, so you are not sent unverified scanner output. Critical issues are raised as soon as we confirm them rather than held for the report.
-
Report & walkthrough
You get an executive summary your board can read and technical detail your engineers can act on — including reproduction steps, what we tested that held, and what was out of scope. We walk your team through it.
-
Retest
Once you have fixed the findings we retest them and confirm in writing what is closed. Retest scope and window are agreed in your scope document before the engagement begins.
Toolkit
- Garak
- TextAttack
- Custom Fuzzers
- LangChain Analysis tools