Prompt injection
Direct messages and indirect input from documents or tool responses. Check whether manipulated instructions reach private data or privileged actions.
Services / AI & LLM applications
AI agent & LLM security testing
Understand what your agent can access, what it can do, and whether untrusted input can push it past those boundaries. Independent manual testing for startups moving AI products into customers’ hands.
Discuss your scope in a 15-min callTest the application around the model.
An assistant can produce a convincing answer while its retrieval layer exposes another customer’s files or its tools act with too much authority. The review follows input from the user and retrieved content through to the data and actions it can reach.
We agree the application, roles, data sources and workflows in scope first. Testing focuses on reproducible application impact, with the context your engineers need to fix it.
What an AI security review can cover
RAG and private retrieval
Whether the retrieval pipeline applies the caller’s permissions before private context reaches the model, including separation between customers and roles.
Agent tools and MCP integrations
Which tools the agent can call, how arguments are checked, and where identity and authorization are enforced around connected services.
Output, credentials and data flow
How responses, browser storage, logs and downstream integrations handle sensitive data. Check the paths beyond the chat window as well as the answer itself.
From a finding to a fix your team can use
- Agree the scope. Identify the agent’s role, integrations, test users and expected permissions.
- Exercise real workflows. Investigate the access and action boundaries with controlled test data.
- Report confirmed impact. Document evidence, reproduction steps, severity, limitations and recommended fixes.
- Check the agreed fixes. One retest is included in the one-time review.
Read the illustrative sample report to see how findings and limitations are communicated.
Evidence of the approach
The Xoscar control-plane research documents a credited report about an internal service boundary. The LLM lab notes show five intentionally vulnerable lab reproductions. The agent tool case study traces access-control and injection failures through a tool implementation.
These examples are public research and lab work, not customer case studies.
Pricing and preparation
The Security Review is a manual penetration test of one agreed application or AI workflow, priced at $2,000, with documented findings, remediation guidance, a summary letter, a results call and one retest; my first three clients pay a founding rate of $1,500. Review + Fix starts at $3,000 and adds implementation of specifically agreed fixes; the final price depends on remediation scope. Ongoing Security Support is $1,000 per month after an initial review, with up to eight hours monthly across selected change reviews, targeted testing, fix verification and consultation. A $500 scoping session is credited toward a Security Review booked within 30 days. Larger work is quoted separately. Compare all plans.
Bring a short product overview, the agent’s integrations, the roles you support and any deadline. We agree the testing window and access arrangements before work begins. If a customer requires a particular report or methodology, bring that requirement so we can check fit.
For the surrounding application, see SaaS and API security reviews. The homepage’s buyer questions cover timing, retesting and customer requirements.
Your next step
Talk through your product.
Bring your product, your timeline, and the questions your customers are asking. We’ll use a 15-minute call to discuss fit and scope.