SilentGrid researches and experiments with AI to establish where it can aid or automate part of an assessment. As AI changes offensive security work, decisions about which tools to adopt require engineering judgement.
Research tests whether a technique improves coverage, holds up against a client's environment and produces results that can be explained afterwards. Unexplained results cannot be used as evidence.
Safe and controlled before useful
Any implementation reaching client work has to be safe and controlled first. Its data handling must be explicit: where data goes, how long it is retained, which jurisdiction it sits in and whether anyone uses it for training. Client data does not enter a tool until those questions are answered.
A consultant remains accountable for the result and verifies each finding before it is reported. Where AI assistance contributes to an assessment, the report describes its role and states the limits of the conclusion with the finding.
The basics still apply, especially now
Most weaknesses found around AI systems are familiar problems introduced through a new component. Established recommendations still apply.
An agent granted broad permissions creates a privilege problem; least privilege applies unchanged. Credentials in a prompt, a system message or a tool configuration are secrets in code, and belong in a secret store. A model gateway exposed to the internet is an internet-facing application that needs authentication, rate limiting and logging like any other. Content in a vector store is data, subject to the same classification and access control as the source it came from. Retrieval that pulls from a share nobody has reviewed inherits whatever is in that share.
Organisations adopting AI quickly are often asked about prompt injection before patching, multi-factor authentication or segmentation. Those established controls still need attention alongside prompt injection; they are usually where the exposure lies.