The evidence layer
for prompt‑injection
testing.
PromptShield runs 217 OWASP LLM Top 10 attacks against your endpoint on every commit and produces a severity‑scored, signed PDF your auditor will accept. Continuous. Reproducible. Paid plans launching in private beta — join the waitlist.
$ npx promptshield scan \
--endpoint https://api.acme.com/v1/chat \
--prompt "Summarize this doc: {doc}"
✓ PSI-2026-0042 CRITICAL Tool-call hijacking via retrieved doc
✗ PSI-2026-0041 HIGH Indirect injection in system prompt
Running 217 attacks... done in 83s
Report: ./promptshield-report-2026-04-25.pdf Built like a security advisory, not a SaaS funnel.
Every attack in PromptShield is documented, reproducible, and severity-scored against the OWASP LLM Top 10. No FUD, no theatre — just the evidence your auditor and your engineering lead both ask for.
Reproducible payloads
Every finding ships with the exact prompt, the response, the CVSS score, and a one-click rerun. Diff against your last green build.
Read →9 of 10 OWASP LLM categories
217 attacks across direct + indirect injection, tool-call hijacking, exfiltration, jailbreaks and prompt leakage. LLM08 (Vector & Embedding Weaknesses) shipping 2026.05. Updated monthly.
Read →Procurement-ready PDFs
Signed, branded, severity-scored. The artefact your customer's procurement reviewer will actually accept on the first request.
Read →CI gate, not a screenshot
One-line GitHub Action or `npx promptshield scan`. Fails the build on regressions; opens a PR comment with the offending payload.
Read →Attack research, not vibes
Every catalogue entry references public incidents, CVE filings or peer-reviewed disclosures. We cite our sources.
Read →Continuous, not point-in-time
Daily catalogue diffs, scheduled rescans, and alert-on-new-attack. Your audit evidence is a feed, not a screenshot from last quarter.
Read →Field notes from the catalogue.
Public incident write-ups, attack technique deep-dives, and the engineering choices behind PromptShield's catalogue. Written for AppSec engineers and ML platform leads, not buyers.
Five public agent-exfiltration incidents and the payloads behind them.
14 min read →Why function-call schemas are the new SQL injection.
9 min read →What an ISO 42001 auditor actually wants to see for prompt-injection controls.
11 min read →
Point us at a staging endpoint.
Get a finding back in ninety seconds.
Paste an API URL and a sample prompt. We'll run five high-signal injection attacks and email you a teaser report. If we find something — and we usually do — the email unlocks the full 25-attack suite.