Every step your agent takes, checked against its task
Every step your agent takes, checked against its task
Darkhunt checks every step an agent takes, before it runs and after: the prompt, each tool call, what comes back and the response. Each step is judged against what the agent was asked to do and what has happened so far. Steps that fit go ahead, and every decision is recorded.
Darkhunt checks every step an agent takes, before it runs and after: the prompt, each tool call, what comes back and the response. Each step is judged against what the agent was asked to do and what has happened so far. Steps that fit go ahead, and every decision is recorded.
What Darkhunt looks at
What Darkhunt looks at
Firewalls and access controls check each action on its own, against fixed rules. Rules miss what nobody thought to write down. Before each step runs and after, Darkhunt looks at six things.
Firewalls and access controls check each action on its own, against fixed rules. Rules miss what nobody thought to write down. Before each step runs and after, Darkhunt looks at six things.
The task
The task
What the agent was asked to do, so an action can be permitted and still not fit.
What the agent was asked to do, so an action can be permitted and still not fit.
The conversation so far
The conversation so far
The earlier prompts, actions and results, so a problem that builds over several steps still shows up.
The earlier prompts, actions and results, so a problem that builds over several steps still shows up.
What the action touches
What the action touches
Which data is involved, and how sensitive it is.
Which data is involved, and how sensitive it is.
What’s normal for this agent
What’s normal for this agent
A baseline of how the agent usually works, so it stands out when it starts acting unlike itself.
A baseline of how the agent usually works, so it stands out when it starts acting unlike itself.
Malicious intent and instructions it shouldn’t follow
Malicious intent and instructions it shouldn’t follow
Signs that a request or input is trying to misuse the agent. That includes instructions that pull it away from its task and hidden text in a document, email or web page, like the line in the invoice.
Signs that a request or input is trying to misuse the agent. That includes instructions that pull it away from its task and hidden text in a document, email or web page, like the line in the invoice.
Your policy
Your policy
The rules you set for what each agent must never do.
The rules you set for what each agent must never do.
Three possible decisions
Three possible decisions
Each step gets one of three decisions, based on the task and the conversation so far. Most go ahead without anyone noticing. Your rules set what happens to a step that doesn’t fit.
Each step gets one of three decisions, based on the task and the conversation so far. Most go ahead without anyone noticing. Your rules set what happens to a step that doesn’t fit.
Allowed
It fits the task
It fits the task
The step goes ahead and is recorded.
The step goes ahead and is recorded.
Flagged
Unusual, but not stopped
Unusual, but not stopped
The step goes ahead, is marked in the record and counts toward what Darkhunt decides next.
The step goes ahead, is marked in the record and counts toward what Darkhunt decides next.
Stopped
It doesn’t fit
It doesn’t fit
Depending on your rules, the step is stopped automatically or held for a person to review. A held step doesn’t run until they approve it.
Depending on your rules, the step is stopped automatically or held for a person to review. A held step doesn’t run until they approve it.
In the claims example, reading extra clinical notes is flagged and sending them to the insurer is stopped.
In the claims example, reading extra clinical notes is flagged and sending them to the insurer is stopped.
Every decision goes into the record, with what the agent did and why Darkhunt allowed, flagged or stopped it.
Every decision goes into the record, with what the agent did and why Darkhunt allowed, flagged or stopped it.
What your team does
What your team does
Start with one agent you want to put into production.
Start with one agent you want to put into production.
Plug in the agent
Plug in the agent
Connect Darkhunt by SDK, gateway or plugin.
Connect Darkhunt by SDK, gateway or plugin.
Say what must never happen
Say what must never happen
Set the rules for what each agent must never do and how a step that doesn’t fit is handled. Darkhunt weighs them alongside the other signals.
Set the rules for what each agent must never do and how a step that doesn’t fit is handled. Darkhunt weighs them alongside the other signals.
Handle only the exceptions
Handle only the exceptions
Review the steps your rules hold for a person. Everything else goes ahead without waiting.
Review the steps your rules hold for a person. Everything else goes ahead without waiting.
Plugs in where your agents run
Plugs in where your agents run
It works with the coding agents, models and agent frameworks below. Run it hosted or in your own environment.
It works with the coding agents, models and agent frameworks below. Run it hosted or in your own environment.
Plugs in by SDK, gateway or plugin
Claude Code
Claude Code
Codex
Codex
OpenClaw
OpenClaw
Anthropic
Anthropic
Gemini
Gemini
SDKs for Python and TypeScript
LangChain / LangGraph
OpenAI Agents SDK
Vercel AI SDK