NewReadiness Kit v0.1 — our open-source agent attack pack and evaluation harness — ships on GitHub on 18 October 2026.Readiness Kit v0.1 — 18 Oct 2026.Open source →
Rainkernel Lab
A top place for an independent lab comes from public work anyone can check, not from hours or claims. The Lab is where Rainkernel does that work in the open: one checkable item every week, security findings reported under our name, benchmarks with published methods, participation in the standards our reports map to, and an evidence pack for every domain we test in.
Where a top place is still open
From a 3 October 2026 review of about 35 AI testing and assurance providers. The certificate for AI vendors, ISO/IEC 42001 certification and red-teaming for the model makers are taken. These are not.
The certificate for AI vendors (AIUC-1), ISO/IEC 42001 certification and red-teaming for the model makers are taken. None of the providers we reviewed publishes a fixed price for testing one deployed agent before launch. The Agent Readiness Gate does: $7,500, two weeks, scope on the page.
Open — we hold it from launch · Agent Readiness Gate →
MCP scanners are free and vendor-backed, but every one scans point-in-time and none publishes a neutral, named assessment of the servers and skills agents actually load. MCPSentry's public findings log and continuous attestation are built for that place.
Open — scanner Q1 2027, first public findings from the Readiness Kit corpus in October · MCPSentry →
Of the roughly 35 testing and assurance providers reviewed, one names runaway cost as something it tests. The Gate measures cost per completed task on real traffic and the Agent Cost Governor enforces the ceiling — a reliability result, not a FinOps afterthought.
Open — measured in every Gate; enforced from December 2026 · Agent Cost Governor →
Rules push buyers hardest in insurance, banking, hiring, manufacturing, pharma and the public sector, and each asks for different evidence. The same test engine with a pack per domain — the questions, the attack cases and the records that regulator expects — is a position no generalist tool holds.
Open — packs published domain by domain from the first sprint in each · Compliance Evidence Pack →
Public agent benchmarks are overwhelmingly English — the multilingual OmnilingualGAIA2 benchmark only appeared in August 2026 — and agents serving customers in Hindi, Telugu, Tamil, Spanish, German or Japanese are shipped on English test results.
Open — multilingual test set on the Readiness Kit roadmap, Q1 2027
Support, coding and browsing have public agent benchmarks; claims triage, prior authorisation, invoice exceptions and customs documents do not. The first public, reproducible benchmark in one of these domains sets the reference others are measured against.
Open — first candidate: a public support-ticket set with the Readiness Kit v0.1, 18 October 2026
How we get there
Flaws found in public agent tools, MCP servers and skills are reported through the maintainers' security advisories and, where they qualify, credited CVEs. The advisory log below lists every one we publish, with the maintainer's acknowledgement.
Our attack pack is mapped to the OWASP Top 10 for Agentic Applications; we contribute cases and findings back to the OWASP GenAI project and take part in the Model Context Protocol registry and security discussions, where the signing and provenance gaps our products address are being worked out.
The AI Verify Foundation runs a Global AI Assurance Sandbox pairing testers with deployers, and is establishing an accreditation programme for AI testing firms — the first government-backed route to being a recognised tester. Rainkernel is applying to both.
Every week the Lab publishes one thing a buyer can verify without talking to us: a benchmark run, a finding, a sample report section, a corpus release, a reproduction of a published agent failure. The log is the record.
How the lab's time is split
Weekly log
Dated, with a link once it exists. A planned item that slips stays on the list as slipped; nothing is back-dated.
| Week | Item | Status |
|---|---|---|
| 3 Oct 2026 | Sample Evidence Report published (fictional claims-triage agent) — the format every Gate delivers | published |
| 12–18 Oct 2026 | Readiness Kit v0.1 on GitHub: attack pack mapped to OWASP ASI01–ASI10, harness seeding, cost baseline, BOM emitter; benchmark run on a public support-ticket set | planned |
| 19–25 Oct 2026 | First teardown: a published agent failure reproduced with the corpus, with the control that catches it | planned |
| 26 Oct–1 Nov 2026 | Permission-audit checklist published; dry run of the Full Gate on a synthetic multi-agent estate | planned |
Advisories
Each entry is added only after the maintainer has acknowledged it. The list starts empty because we have published none yet; the first is expected in October 2026 from the Readiness Kit corpus work.
No advisories published yet. Reports go to maintainers first; the entry appears here when they acknowledge it, with the advisory ID and, where issued, the CVE.
Found something in one of our tools? security@rainkernel.com · security.txt. We credit reporters.
Domain packs
The engine — attack pack, harness, cost meter, permission audit, BOM, recorder — is the same everywhere. What changes is the regulator's question, the attack cases that matter and the record that has to exist. Each pack is published after the first Gate in that domain.
Rules push hardest
What makes buyers prove it. State AI bulletins adopted across most US states; Colorado's rules on unfair discrimination in insurance; CMS-0057-F prior-authorisation APIs for US payers from January 2027; IRDAI outsourcing and governance expectations in India; EU AI Act high-risk pricing and claims categories.
Pack output. Decision record with explanation and human-review trail; bias-probe results; model and prompt version inventory
Rules push hard
What makes buyers prove it. Model-risk management expectations (SR 11-7 in the US, RBI's FREE-AI framework in India); EU AI Act high-risk credit scoring; AML and KYC record-keeping; consumer-protection rules on automated decisions.
Pack output. Model-risk validation pack: evaluation set, challenger results, monitoring plan, change records
Rules push hard
What makes buyers prove it. New York City Local Law 144 bias audits; Illinois's AI-in-employment amendment from 1 January 2026; Colorado SB 26-189 from 1 January 2027; the EU AI Act's employment high-risk category (December 2027).
Pack output. Bias-audit report in the published format; notice and explanation records; evaluation set and results
Rules push hard
What makes buyers prove it. The EU Machinery Regulation from January 2027 (AI in safety functions); the Cyber Resilience Act for products with digital elements (reporting from September 2026); ISO 9001 change control; functional-safety standards.
Pack output. Agent BOM; change and incident records; authority-limit test results
Rules push hard
What makes buyers prove it. GxP computerised-system validation (EU Annex 11, 21 CFR Part 11); FDA and EMA guidance on AI in regulatory decision-making; pharmacovigilance reporting timelines.
Pack output. Validation pack (IQ/OQ/PQ-style evidence for the agent), tamper-evident audit trail, periodic review records
Rules push hard
What makes buyers prove it. Public-procurement rules asking for inventories and impact assessments; records and freedom-of-information laws; the EU AI Act's public-service high-risk category; US federal and state AI-use policies.
Pack output. AI inventory entry, impact-assessment inputs, decision records, accessibility results
Rules push hard
What makes buyers prove it. ONC HTI-1 transparency for decision-support software; state laws on generative AI in patient communications; HIPAA; EU MDR where software is a medical device.
Pack output. PHI data-flow map, evaluation results a clinical reviewer signs, disclosure records
Rules push moderately
What makes buyers prove it. EU AI Act transparency duties (Article 50, from August 2026); the FCC's ruling that AI voices in calls fall under the TCPA; client contracts' data-residency and audit clauses; SOC 1 and 2 for service organisations.
Pack output. Containment and resolution QA score, surge test report, cost per resolution
Rules push moderately
What makes buyers prove it. Professional-conduct and confidentiality rules; court orders on AI-assisted filings; client audit clauses; ISO/IEC 42001 adoption among firms.
Pack output. Grounding and citation audit; confidentiality test results; usage records
Rules push moderately
What makes buyers prove it. SOC 2 and ISO 27001 questionnaires with AI sections; the Cyber Resilience Act for products sold in the EU; EU AI Act GPAI and transparency duties; customers' procurement asking for an agent inventory.
Pack output. Agent BOM for the security questionnaire; attack-pack results; cost per task
Rules push least
What makes buyers prove it. Customs and trade documentation rules; SOX controls around purchase-to-pay; contractual SLAs with carriers and customers — little AI-specific regulation.
Pack output. Cost-per-document baseline, exception-handling evaluation, change records
Rules push least
What makes buyers prove it. Consumer-protection and advertising rules; PCI DSS where agents touch payments; DPDP and GDPR for customer data — little AI-specific regulation.
Pack output. Action-limit test results, surge report, cost per resolution
Languages: every pack runs in English and in the languages the agent serves customers in; the multilingual test set is on the Readiness Kit roadmap for Q1 2027. Sector detail on the industries page; the evidence exports on the Compliance Evidence Pack page.
Design partners get the Gate at cost in exchange for a published, anonymised case. Integrators get the toolkit. Either way the result is a report anyone can check.