The Hacker News article, contributed by a Pentera product marketing executive, argues that the engineering in agentic penetration testing sits in the harness around the AI model, not in the model itself. A general-purpose model can write attacks, but every request must be checked against plain-language scope rules, and privileged actions should pass through a deterministic policy engine at execution time, per OWASP's 2026 guidance.
Prompt injection is called the top LLM application risk, with no reliable prevention available today. Because any page or API response the agent reads can carry malicious instructions, the article says the real question is how tightly each action is constrained and which actions require human sign-off. It also warns that approval fatigue can degrade reviewer judgment, so routing everything to a human merely shifts the failure point.
The piece cites Google's October 2026 decision to stop accepting product vulnerability reports to its open-source bug bounty program after a surge of mostly invalid automated submissions. That reinforces the need for proof: a finding must come with a replayable exploit, a CVSS score, and enough detail for a developer to reproduce it. The article then describes Pentera Apps as implementing this separation of reasoning and execution, with logging, pause controls, redacted data, and replayable findings built in. As a contributed vendor piece, it should be read as a product-oriented argument rather than independent research.