Design partner beta
Limited to 10 teams for 4 to 6 weeks. 4 spots remaining.
Teams running AI agents that call tools or touch sensitive data, with a named security or engineering owner for the pilot.
All four checkpoints on one agent. We start in report-only mode, so nothing is blocked until you decide. Direct access to the engineers.
A named technical contact, weekly feedback, and your verdicts on the detections we get right and wrong.
No commitment. We review every application and reply by email. Spots go to the best fit, not first come. Data handling terms are shared before you connect anything.
For security teams
No. AgentHacker is added with SDKs for Python and TypeScript. There is no gateway to route through and no rewrite of your agent, and it works with any model.
Policy is enforced at runtime, and tool calls are authorized at checkpoint CP-3. Authorization is hash-bound: a call is locked to the record’s owner and to its exact arguments, so it cannot be pointed at a different record or altered after approval.
Every decision is written to an append-only audit ledger. Entries are signed in sequence, and administrators cannot edit them.
Alerts go to Slack and can be routed to the right owner: the CISO, a security engineer, or the compliance lead.
No. The beta starts in report-only mode, so nothing is blocked until you decide to turn enforcement on.
Data handling terms are shared before you connect anything, so your security team can review them first.
On two public benchmarks, AgentDojo and AgentDyn, 0 of 642 attacks succeeded with AgentHacker in the loop, against 221 of 642 with no defense. It interrupted 0 of 62 benign episodes. Method and full tables are on the benchmark page.
Beta program
AgentHacker is a runtime security platform that detects and prevents prompt injection and adversarial attacks against autonomous AI agents. Four checkpoints—input, tool-response data, tool calls, and output—inspect every step of the agent loop in real time, catching attacks that other solutions miss.
AgentHacker installs as a framework-agnostic SDK inside your agent loop. You'll wrap your agent's run function with our SDK, which logs all activity and flags threats in real time. We'll provide SDKs, API docs, and a dedicated Slack channel. Most integrations take a few hours.
Either works. Staging is great for testing edge cases and validating our detections. Production traffic gives us more real-world adversarial patterns. We recommend starting in staging for 1–2 weeks to validate the integration, then moving to production.
Yes. We support self-hosted and on-prem deployments for regulated workloads. If your data can't leave your environment—financial services, healthcare, government—we have a deployment model that works.
We'll sign an NDA covering any agent logs or system details you share. You can redact sensitive information before sending logs to us. All data is encrypted in transit and at rest, and deleted at the end of the beta unless you explicitly opt in.
You'll have three options: (1) Continue with AgentHacker at the discounted beta pricing, locked in through our first year of GA; (2) Upgrade to a paid plan with additional post-GA features; or (3) wind down. Your feedback directly shapes the product.
Yes. We provide detailed logs, audit trails, and compliance reports (SOC 2 coming). Many beta customers are using AgentHacker to satisfy security requirements from their boards, customers, or regulators.
That's the whole point. Report them and we'll prioritize fixes based on severity and your timeline. Critical issues get 24–48 hour turnaround.
No hard lock-in. The beta is designed to be a 4–6 week trial. After that, you can choose to continue, upgrade, or step back.
Fill out the form or book a 30-minute intro call. We'll get you set up with SDK access and documentation within 48 hours.
Prefer to talk first? Book a 30-minute intro call. We will answer technical questions and walk through a demo.
Book 30-min call