AgentHacker sits inside the agent loop and stops attacks before the agent acts on them. This page is the short version of why we think that is a company.
What we have shown
Two public benchmarks, one fixed victim model, competitors on their published configurations. Our failures count against us.
attacks succeeded against an agent running AgentHacker.
benign episodes with an intervention, so the agent keeps doing its job.
Why now
An agent that reads a web page, an email or an API response will follow instructions hidden in it. Prompt injection is the first item on OWASP's Top 10 for LLM applications.
Large security vendors have been acquiring AI security startups since 2024. Forrester says acquirers are paying for talent, early Fortune 500 traction, and a story they were missing.
In our benchmark the open-source options either let attacks through or broke normal work: LlamaFirewall let 29.0% of attacks succeed and intervened on 38.7% of benign episodes; NeMo Guardrails intervened on 54.8%.
Where we sit
A guardrail can stop attacks and still block the agent's normal work. Buyers need both numbers, so we publish both. Here is how the main approaches compare.
Category
Prompt and model guardrails
What it does
Screen prompts and replies for attacks and policy violations.
What it leaves open
Typically built around the prompt and the reply, not the agent's tool calls.
Category
Discovery, posture and governance
What it does
Find the agents and AI tools in use, then set and monitor policy.
What it leaves open
Shows where the risk is. Stopping it inside the agent loop is a separate job.
Category
Lifecycle platforms
What it does
Discover agents, test them before launch, monitor them in production.
What it leaves open
Broad scope. Buyers still need proof that normal work is left alone.
Category
Sandboxes and action checks
What it does
Limit what an agent can touch, or verify an action against policy before it runs.
What it leaves open
Narrow by design, so usually paired with detection.
Category
AgentHacker
How
Four checkpoints inside the agent loop: inputs, tool data, tool calls, outputs.
Evidence
Tested head to head against two named competitors on public datasets. Figures are in the benchmark report.
Categories overlap and the descriptions are general. Benchmark figures are ours; see the report for method and limits.
Why it compounds
Wrapping a model is not a moat. These are the three things we are building the company on. The tags show what is live today and what is still to come.
We sit between the model and its tools, so we see the whole trajectory, not only a prompt. That is what lets us stop an attack at the tool call.
Every verdict on a detection is training signal. Detectors trained on design partner signal are the next build. They do not exist yet.
Public datasets, competitor configurations and our own misses are all in the report. A buyer can check the claim instead of taking our word.
Where we are
Talk to us
Write to Haider Shareef, CEO. hshareef@agenthacker.ai