Sensitive actions

Stop risky AI actions before they become security incidents.

A safe prompt does not guarantee a safe execution path. Once an agent begins working, its inputs are files, dependencies, tool responses and fetched content — none of which the original request was checked against. Oconee Runtime detects the risky patterns that emerge at that point, such as credential exposure and destructive commands, and evaluates the resulting action against policy.

No sales call required.

Prompt safety and action authorisation are different controls

Prompt filtering answers a question about the input. It is useful and it is not sufficient, because an agent's execution path is assembled from material the filter never saw: a file it opened, a dependency it resolved, a tool response it trusted, a page it fetched.

By the time the risk is real, the relevant object is no longer a prompt. It is a proposed action against a specific resource, and that is what has to be authorised.

Example scenario

One action, evaluated

Proposed action

Coding agent

Command referencing a credential file

Context

Actor
Developer
Agent
Coding agent
Resource
Credential material
Action
Command execution
Signals
credential_leak, sensitive_file_reference

Policy decision

BLOCK

Credential signals resolve strictly. The decision and the signals that produced it are recorded, so the response is an investigation rather than a guess.

Recorded as evidence

  • actor
  • agent
  • action
  • resource
  • context
  • policy
  • decision
  • timestamp
Examples illustrate Oconee Runtime workflows and policy scenarios.

The signals behind the decision

Oconee names what it detected rather than returning an undifferentiated risk score. A named signal is checkable against the product and actionable in a policy; a score is neither.

  • Credentials — credential_leak, api_key_exposure, private_key_exposure, session_token_exposure
  • Commands — command_execution_intent, destructive_command_intent
  • Files and dependencies — sensitive_file_reference, dependency_risk_detected
  • Untrusted input — prompt_injection_attempt, untrusted_source_detected, unsafe_instruction_detected
  • Data exposure — pii_detected, source_code_exposure, secret_in_url
  1. Actor
  2. Agent
  3. Action
  4. Resource
  5. Context
  6. Policy
  7. Decision
  8. Evidence

Relevant capabilities

  • Named signal detection across supported browser and engineering surfaces
  • Allow, warn, block and redact outcomes
  • Per-signal enforcement posture, so detection can precede blocking
  • Decision evidence recorded with the signals that produced it

Detection before enforcement

Signals can run in a monitoring posture before anything is blocked, which is usually the right sequence: see what the organisation actually does, then decide what to refuse. Enforcement modes beyond monitoring depend on plan.

Common questions

Why is prompt filtering not enough?
Because an agent's risk is created after the prompt, from files, dependencies, tool responses and fetched content the filter never inspected. The object that needs authorising is the resulting action, not the request that started it.
How do you control AI agent actions?
By evaluating the proposed action against the conditions present when it is attempted — actor, agent, action, resource, context and policy — and resolving it to allow, warn, block or redact, with the decision recorded.

AI proposes. Policy decides.

All use cases