Prompt injection
Attempts to override instructions through direct user prompts or indirect content retrieved into the model context.
Guide
AI red teaming tests behavior, not just code.
LLM and agentic systems are probabilistic and context-dependent. Pentoma AI Red Teaming focuses on adversarial behavior, replayable prompts, and evidence that product teams can act on.
Core concepts
The attack surface includes prompts, context, tools, and agent scope.
AI red teaming is the clearest public language for this work because it communicates adversarial testing without pretending LLM failures behave like traditional vulnerabilities.
Tool misuse
Attempts to make an agent call tools, plugins, or functions outside intended authorization boundaries.
Sensitive disclosure
Attempts to expose secrets, private data, system prompts, or internal context through model behavior.