Secure your LLMs.
Evidence for your compliance file.

Shield LLM automatically sends OWASP LLM Top 10 attack prompts to the target chatbot endpoint. Each response is analyzed by our 3-layer engine (context + rules + LLM-as-Judge). You get an A–F grade, the vulnerability breakdown, and an audit-ready PDF report.

Shield LLM is self-hosted: there is no Shield server. Attack prompts go straight to your chatbot, and the replies are scored by the instance you run on your own infrastructure, which also stores the history shown in your dashboard. The only outbound call is the AI judge, to the LLM provider you configure with your own API key, replies are used solely to score the scan, never to train models.

Yes, as long as it is reachable over HTTP. Point the CLI at your endpoint, whether that is an OpenAI-compatible API, Claude, Gemini, Mistral or any custom chatbot, and shield.config.json handles authentication, headers and request/response mapping. SSE streaming is auto-detected. No model-specific configuration required.

Mostly it depends on your chatbot, not on Shield: attacks run 4 at a time by default, so your endpoint's own latency and rate limits set the pace. A quick scan takes well under a minute. A full scan, which adds the multi-turn attacks, takes several minutes: we measured just under 6 minutes for the attack run against a public test target, and a rate-limiting endpoint takes longer still because the scanner automatically backs off instead of hammering it. Raise --concurrency to trade politeness for speed.

Weighted Impact Scoring: every finding carries a severity weight (CRITICAL 10, HIGH 5, MEDIUM 2, LOW 1), and the score is 100 × (1 − observed penalty / total weight), so one critical finding costs far more than a handful of low ones. Grades map to A (≥90), B (≥75), C (≥60), D (≥40), F (<40). The calculation details are documented in the PDF report.

It covers part of it, and the report says which part. Article 50 transparency is the obligation that binds a chatbot today, and Shield tests it by checking whether your system claims to be human. It cannot see a disclosure banner in your interface, so it never scores that as a breach. Articles 9, 10, 13, 14 and 15 are scored from the scan, and the report marks them as good practice rather than obligations when your system is not high-risk. Organizational articles come from a questionnaire and are self-declared. Nothing here is a conformity assessment or a certification, and the legal call stays with your DPO.