superagent_

// models

Examples

[ view markdown ]

End-to-end Security-One pipelines that decide clear cases immediately and escalate uncertain or risky ones to a stronger OpenAI model.

Each example runs Security-One as the always-on first pass. It decides clear cases from calibrated probabilities and sends only uncertain or risky cases to a stronger OpenAI model, so you pay for deep analysis where it changes the outcome.

flowchart TD
  A[Event] --> B[Security-One]
  B -->|"clear"| C[Act in your code]
  B -->|"uncertain or risky"| D[Stronger OpenAI model]
  D --> C
  D --> E[Human review]
Example Security-One decides Escalates when Result
Detect prompt injection Whether untrusted input is a prompt-injection attempt The probability falls between 0.30 and 0.70 Allow, block, or send to human review before the input reaches your agent
Check pull requests Security risk, severity, and area for each changed file Risk is 0.30 or higher, High or Critical severity is at least 0.50 likely, or confidence is below 0.50 A pull request comment with line-level findings and a failing check for high or critical issues

What every example shares

  • Thresholds and actions live in your code. Security-One returns probabilities; your code decides what each band means.
  • The stronger model gets structured input and output. It receives the original content plus Security-One's answers and must reply in a fixed JSON schema.
  • Untrusted content stays data. The stronger model reads the same untrusted content, so it gets no tools and is told never to follow instructions inside it.
  • Failures fail closed. In the prompt-injection example, a failed escalation sends the input to human review and a failed Security-One call throws before the input reaches your agent. In the PR check, any model failure or unanalyzed file fails the check.

Next steps