01 //
Indirect prompt injection
A hostile instruction arrives inside a retrieved document and the agent treats it as an order.
// Agent audit
The agent era is here. Most of those agents are leaking data.
An assessment of one RAG, tool-using or agent pipeline. The report lists indirect injection, PII leakage and out-of-policy tool calls.
$ deepwave-audit --pipeline agent-rag --simulate-jailbreak
[RISK] retrieved document carries an indirect instruction
[CRITICAL] tool call outside policy
[CRITICAL] PII pattern in the context sent to the model
› 2 findings · assessment recommended
// What the assessment covers
Six failures the report marks
02 //
PII exfiltration
Personal data enters the model context or leaves in the answer and the tool log.
03 //
Excessive tool execution
The agent calls a tool beyond what the task policy allows.
04 //
Cost of hallucinated loops
The agent repeats a model or tool call with no stop condition.
05 //
LGPD/GDPR control gap
Legal basis, retention and access logs do not cover what the pipeline processes.
06 //
Context poisoning
Content stored in memory or in the index starts steering later tasks.
// How to engage
Two ways in
01 // R$ 297
Technical playbook + Docker boilerplate
A control checklist and a sample compose with sanitization and a circuit breaker. This site does not charge. The request goes through this form and the payment link comes by email.
Request the playbook →02 // By proposal
5-day security assessment and red team
One production or staging pipeline. The delivery is the finding list with severity and the prioritized fix. No maturity package.
Request a proposal →