Pentera says the harness, not the model, makes agentic pentesting safe
Original title: The Harness Is the Hard Part of Agentic Pentesting
Pentera Apps pairs a model that chooses tests with code that logs, pauses, and stops every action, and ships replayable exploit scripts.
A Pentera product-marketing piece argues that safe agentic web pentesting depends on a deterministic harness around a general-purpose model. It cites scope enforcement, credential policy, prompt injection, and reproducible proof, and points to OpenAI’s July disclosure of a Hugging Face intrusion and Google’s October 2026 halt of automated open-source bounty reports. Pentera Apps is described as logging every action, limiting data sent to the model, and attaching a replayable script, CVSS 3.1 score, and CWE mapping to each finding.
Unconfirmed: This is a contributed Pentera marketing piece; the Hugging Face intrusion is only referenced, not newly documented.
- The piece is written by Assaf Regev, AVP of product marketing at Pentera.
- Google stopped taking product reports to its open-source bounty after mostly invalid automated submissions.
- Pentera says only minimal redacted data reaches the LLM, with zero retention and no training on customer data.
- Change-triggered testing in CI/CD is described as coming next, not as already available.
- Vendors
- Pentera
- Products
- Pentera Apps
- Organizations
- OpenAIGoogleOWASPHugging Face
From the original article
4 min read at thehackernews.com
An AppSec team can point an AI agent running on a general-purpose model at its own web apps with little setup. Sure, the model writes the attacks. But everything that makes those attacks safe to run against production and worth acting on afterward lives in the harness around the model.
thehackernews.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.