Loading…...
A scoped pilot for teams shipping an LLM application who need a defensible answer to: which defenses actually hold, which ones fail, and what to fix first. Authorized scope, human-reviewed remediation, replay-validated retest, sealed evidence pack.
Every stage produces an artifact. Nothing leaves scope.
You tell us which LLM application is in scope, the endpoints or surfaces to evaluate, and what success looks like for your team.
Engagement brief · asset list · success criteria
We require a signed Rules of Engagement and proof that you own or control the target. DNS TXT challenge is the default ownership proof.
Signed ROE · DNS-verified target ownership · named approval owner
The engine drives a version-locked library of OWASP LLM Top 10 probes and adversarial techniques against your authorized scope. Probes respect the scope boundaries you approved.
Per-technique transcripts · probe inventory · SHA-256 evidence hashes
Every bypass is traced back to the specific control that should have stopped it — refusal filter, output validation, rate limit, tool-access auth, secret scrubbing — with severity and reproducible evidence.
Findings ranked by severity · vendor-specific defense gap · MITRE ATT&CK mapping
For each finding we propose a system-prompt patch, a guardrail config, or a control change. Your team reviews and approves every change before it is replayed.
Proposed patch · guardrail config · reviewer sign-off
We replay the same attack against the patched target. Findings that close on replay are marked validated; findings that still bypass move into the next cycle.
Replay diff · validated vs. still-open findings · retest coverage report
The engagement closes with a sealed evidence pack: SARIF, CSV, JSON, conversation transcripts, remediation playbook, and checksum manifest — ready for an auditor or board.
Signed evidence pack · per-framework SARIF · remediation playbook
No benchmarks are claimed. The pilot reports only what was observed during this engagement.
Elapsed time from authorization to the first confirmed, reproducible finding.
Findings traced back to a specific control, with reproducible evidence and a proposed remediation.
Elapsed time from a confirmed finding to a reviewed, approved, and replay-validated closure.
Share of confirmed findings replayed after remediation, including validated and still-open outcomes.
Request pilot access and we will reply within two business days with a draft Rules of Engagement and target ownership instructions. No anonymous targets — proof of ownership is required before any probe is sent.
Pilot engagements are operated under signed DPA and ROE. Customer data and findings are tenant-isolated and SHA-256 checksummed.