We break your AI before the adversaries do.
A scoped evaluation that starts with a 91-case project regression corpus, adds tests for your system, and reports what the run actually finds. Indicative engagements: $5K–50K.
If you have deployed an LLM product -- a chatbot, an agent, or a RAG system -- you need evidence about how its actual integration behaves under adversarial input. The useful result is a reproducible finding tied to your system, not a generic security headline.
SCBE-AETHERMOORE includes a 91-case project-authored regression corpus organized by adversarial tier. Its 91/91 in-project result measures corpus fit, not transfer to a client application. A scoped engagement adds relevant cases and reports the observed misses, false positives, limitations, and remediation priorities.
What we test for
Four categories, dozens of sub-vectors each. Every finding reproducible, every vector mapped to your application.
Prompt injection
Direct prompt injection, indirect injection via retrieved documents, multi-turn drift attacks, and context poisoning through memory or session state.
Jailbreaks
Persona hijacks, instruction-override attacks, role confusion, safety classifier bypass, and encoding-based evasion techniques.
Data exfiltration
System prompt extraction, training data leakage, PII recovery via inference attacks, and confidential context bleed across sessions.
Agent abuse
Tool misuse, privilege escalation through chained calls, infinite loops, cost bombs, and unauthorized lateral movement across connected systems.
The process
A four-week cycle is an indicative target for a bounded endpoint. The written scope controls the actual timeline.
Scoping
Kickoff call, threat model review, test suite selection. We learn what matters to you and which attack surfaces your architecture exposes.
Execution
Run the agreed project and system-specific cases against the scoped endpoint. Capture each request, response, evaluator decision, and uncertainty in structured output.
Analysis
Findings triaged and severity ranked. Remediation drafted against your stack. We filter the noise so you get the signal that matters.
Delivery
Branded PDF report, live walkthrough call with your engineering and security stakeholders, prioritized remediation roadmap delivered as markdown.
Three indicative scopes.
These bands help start a conversation. The final price, test count, timeline, and deliverables follow the written scope.
- Four-week target after access is ready
- Project corpus plus agreed system-specific cases against one endpoint
- Standard branded PDF report
- Raw JSONL test output
- Remediation roadmap
- For startups and single-product validation
- Everything in Quick Scan
- Custom threat model tailored to your architecture
- Multiple endpoints and agent surfaces
- Executive briefing deck for leadership
- Remediation support during fix cycle
- For scaling companies with multiple AI surfaces
- Everything in Deep Engagement, quarterly
- Drift monitoring between scheduled runs
- Priority response to new attack patterns
- Dedicated Slack channel
- Named point of contact on our side
- For enterprises with ongoing risk
What you get
Every engagement ships with artifacts your engineering, security, and compliance teams can actually use.
Branded PDF report
Your logo, your threat model, your priorities. Written to be read by engineering leadership, not just security specialists.
Severity-ranked findings
Every finding gets a reproduction recipe: exact prompt, exact response, exact conditions. If we found it, you can reproduce it.
Remediation roadmap
Prioritized fix list with code snippets where applicable. Patches, policy middleware, prompt hardening, output filters.
Raw test output (JSONL)
Full structured output for your GRC and audit teams. Feed it to your SIEM, your compliance tooling, or your own dashboards.
Live walkthrough call
Screen-share session with your engineering and security stakeholders. We walk every critical finding and answer questions live.
30 days of follow-up
Email and async support while you implement fixes. We validate your patches against the original attack vectors at no extra charge.
Examples of the findings a scoped test can produce
These are illustrative scenarios, not claims about completed customer engagements. A real report contains only findings reproduced on the system under test.
Read the project-authored evidence, archived comparisons, and limitations →
31 prompt injection vectors, 4 critical
Customer-facing assistant leaked coupon logic and allowed arbitrary refund approvals through indirect injection via product reviews.
policy middleware with strict input/output separation
System prompt extractable via encoding trick
Base64-wrapped payload bypassed the model's refusal training and surfaced the full system prompt including vendor-specific guardrail instructions.
output filter with regex and semantic leakage detection
Training data exfiltration via follow-up questions
Multi-turn questioning reconstructed verbatim chunks from the indexed corpus, including one document marked confidential in metadata.
query classifier + retrieval-time ACL enforcement
The technology underneath
SCBE-AETHERMOORE is the adversarial framework powering every engagement. Open-source core, patent-pending architecture.
The project test taxonomy is organized in six tiers: L1 basic through L6 adversarial. Coverage varies by implementation and corpus; the current public evidence should be read as project-authored testing, not independent certification.
The core is open source. Inspect it, run it yourself, fork it. What we sell is the expertise, the curation, the reporting, and the remediation follow-through.