About this site

modelbouncer.com explains runtime gates for AI models and agents: the guardrails, policy checks and approval steps that sit outside a model and decide which inputs, outputs and actions are allowed through.

Last reviewed: 29 October 2026

What the site covers

• Prompt injection and agent hijacking, and why controls outside the model matter.
• Guardrails on model inputs and outputs.
• Policy enforcement on tool calls, least privilege, and human approval for high-impact actions.
• How these controls relate to public frameworks such as the OWASP Top 10 for LLM Applications, the NIST AI Risk Management Framework and MITRE ATLAS.

“Model bouncer” is used as a descriptive phrase, not an established industry term. The site carries no advertising.

Method

• Primary sources. Pages draw on standards bodies and government agencies (OWASP, NIST, the UK NCSC, MITRE) and on peer-reviewed or preprint research, such as papers on arXiv. Vendor marketing material is not used as a source.
• Every number cited. Each figure and factual claim carries a numbered citation to a source that was read, with the date it was accessed. The source list collects them all.
• Dated reviews. Each page shows a “Last reviewed” date. A review means re-reading the cited sources and re-checking links. Substantive changes to a published page get a dated note on that page.

How the site is made

Automated tools, including AI language models, help research and maintain this site: finding and summarising sources, drafting text, building the pages and checking links. Before a change is published, its claims are checked against the cited sources. A claim that cannot be traced to a source is left out.

Corrections

If you find an error, an outdated figure or a broken link, please send an email with the page address and, if possible, a source. Confirmed errors are corrected on the page, and the “Last reviewed” date is updated.