Red teaming
We test agent behavior with authorized adversarial scenarios and document failures, impact and retesting.
ARKEA IA / SOLUTIONS
Security consulting for AI agents.
ARKEA IA provides security advice and evaluation for chatbots, assistants and agents: red teaming, prompt and context injection testing, bypass assessment and guardrail design. Testing covers authorized systems within an agreed scope.
AI CONSULTING & SECURITY
Before giving it autonomy, test its boundaries.
We advise teams building or already using chatbots, assistants and agents. We review the model, context and tools as parts of one system.
Prompt & context injection
We evaluate malicious instructions in messages, documents and tools, including scope manipulation.
Guardrails & bypass
We assess control bypasses and design least-privilege permissions, action validation and human escalation.
“Is disguising a question enough to bypass a safeguard?”
A reformulation can sometimes expose a weak control; it is not a universal rule. Reproducible tests provide the answer. A system prompt does not replace server-side authorization, separation of data and instructions, or review of sensitive actions.
What do we evaluate?
| Prompt injection | Instructions that attempt to change expected model behavior, directly or through external content. |
|---|---|
| Context injection / scope manipulation | Content that attempts to alter context, objectives or permitted scope. We evaluate source provenance and separation of content from authority. |
| Bypass | Circumvention of a control. It is assessed against defined scenarios and evidence; an unexpected answer alone does not prove unauthorized access. |
| Guardrails | Input, output and execution controls: validation, permissions, limits, logs and approval. They reduce risk; they do not guarantee a model never fails. |
| Destilación / Distillation | Adapting a smaller model using examples or outputs from another model with appropriate usage rights. We assess quality, cost, privacy and security regressions before recommending it. |
What does an engagement deliver?
- Agreed scope: assets, permissions, data, actions and explicit limits of the assessment.
- Reproducible test cases: preconditions, evidence, observed behavior and impact.
- Prioritized findings: affected components, severity and recommended controls.
- Remediation and retesting: comparison against the initial cases and a regression suite for future changes.
When does distillation make sense?
When a well-defined task has suitable examples and a smaller model could meet quality targets with lower latency or cost. It requires a comparison against the current model; it is not an automatic security improvement.
Technical references
Our evaluation vocabulary draws on the OWASP description of prompt injection and the NIST AI Risk Management Framework. These references do not imply certification.