Internal red-team self-assessment (April 2026): 5 high and 2 medium findings, all fixed, each with a regression test. No third-party penetration test yet. Detection scored on a 2,580-sample held-out set it was never trained on: 0.788 recall · 0.857 precision, 3.12% false-positive rate. Hash-chained audit log, with 7-year WORM archival when S3 Object Lock is configured.
When you secure APIs, AI agents, LLM traffic, and deploy plugins — you need to trust the security platform itself
Your Security Tool is an Attack Vector
Every security platform you add to your stack becomes a potential target. If your API security gateway is breached, attackers have full visibility into your APIs and can disable protections.
Black Box Trust Problem
Most security vendors ask you to trust their infrastructure without visibility. You don't know who has access to your logs, how data is encrypted, or if compliance controls actually work.
Vendor Lock-In Risk
Cloud-only security platforms create single points of failure. If the vendor has an outage, your APIs go down. If they change pricing, you're stuck. If they shut down, you lose everything.
Security Testing, Stated Plainly
Internal red-team self-assessment (April 2026): 5 high and 2 medium findings fixed, each with a regression test; no third-party penetration test yet. A CI regression-stability gate runs the detection corpus on every commit with zero false positives on benign inputs — regression fails the build.
Tamper-Evident Audit Trail
Audit log entries written to append-only storage with SHA-256 integrity verification — modification breaks downstream hash verification. 84-month WORM archival when S3 Object Lock is configured. Three verification levels (full chain, single entry, last-N).
Self-Hosting Available · No Lock-In
Run G8KEPR as SaaS or deploy on your own infrastructure via Docker containers and Kubernetes Helm charts. Export all data anytime in standard formats (JSON, CSV). Built on open standards (REST, OAuth, JWT) for portability.
Production-ready security features protecting your APIs and AI agents
AES-256-GCM and ChaCha20-Poly1305 cipher suites. HSTS preload. mTLS between sensors and the collector.
Application-level encryption for API keys, OAuth secrets, TOTP, webhook secrets, MCP secrets. 96-bit nonce per operation, 128-bit auth tag.
PBKDF2-HMAC-SHA256 with 600,000 iterations (OWASP 2024). Passwords use bcrypt at cost factor 12+. Timing-safe comparisons everywhere.
Each entry signed with HMAC-SHA256 + deployment-specific key. Genesis block derived from the key itself. DB access alone cannot forge a valid chain entry.
90 days hot in PostgreSQL (monthly partitions); 7 years cold in S3 with Object Lock COMPLIANCE mode when archival is configured. Self-hosted deployments set their own bucket.
Full-chain verification (every entry), single-entry spot check, and last-N rapid integrity check. Detect any retroactive modification by recomputing hashes.
Every merge gates on CodeQL, Bandit, Trivy, pip-audit, npm audit, OWASP ZAP, Gitleaks, and Semgrep. Build fails on any new finding.
JWT_SECRET rotates every 90 days, ENCRYPTION_KEY every 1 year. Decryption keeps previous key as fallback so rotation is zero-downtime.
Until September 2026 this page described an April 2026 external penetration test by “Pentx Security”. That was mislabelled. The work was an internal red-team self-assessment, and no third-party penetration test has been done yet.
What the internal assessment found is real: 5 high and 2 medium issues, all fixed, each with a regression test in the repository. The three medium findings previously listed here (robots.txt, marketing-subdomain headers, IPv6 rate-limit normalisation) came from the mislabelled report, had no tests behind them, and have been withdrawn.
Retracted: “The MCP Design Flaw Affecting 200,000+ Servers”. That post reported original research — a passive scan of public MCP server configurations, with percentages for how many lacked tool-registration authentication. No such scan was ever run. There is no scanner, no dataset and no methodology behind it, so the post has been unpublished rather than corrected, and the figures should not be cited by anyone who read it. The protocol-level concern it described — that MCP’s flat tool namespace has no cryptographic binding between a tool name and a server identity — is real and is why our own interception path pins tool definitions by hash, but we have no measurement of how widespread the exposure is. We will not publish one until we have data.
Detection is scored on a frozen 2,580-sample held-out set it was never trained on: 0.788 recall, 0.857 precision, 3.12% false-positive rate. The test fails the build on any regression.
Held-out set: 496 attacks and 2,084 benign prompts.
Each STRIDE category mapped to a concrete platform mitigation, not generic hand-waving:
Building towards comprehensive compliance with transparency
2,000+ controls mapped across 12 compliance frameworks with automated assessment, plus 8 reference catalogs. Every audit-log entry, every rate-limit event, every policy decision is mapped to specific control IDs — auditors get exports, not spreadsheets. Maturity language follows the data-room baseline (March 2026).
G8KEPR provides the technical controls and evidence — your auditor issues the certification. The maturity labels above match our internal data-room baseline. Below is what we will and will not say about ourselves until attested by a third party.
What has been tested, by whom, and what has not happened yet
Full report available under NDA. Contact security@g8kepr.com
Public key pinning
The release public key will be published with the first signed pack so customers can pin it. Until then the sensor ships a placeholder key and refuses to load packs signed with it.
Secure your security platform
Internal red-team self-assessment (April 2026): 5 high and 2 medium findings fixed, each with a regression test. No third-party penetration test yet. Findings published, not hidden.
A CI regression-stability gate runs the detection corpus on every commit with zero false positives on benign inputs — any detection regression fails the build. Plus 8 security scanners (CodeQL, Bandit, Trivy, pip-audit, npm audit, OWASP ZAP, Gitleaks, Semgrep) gate every merge.
Enterprise security controls built-in: TLS 1.2+ and AES-256 encryption, RBAC with MFA, append-only audit logs, Ed25519-verified pattern packs, and an internal red-team self-assessment (April 2026).
Common questions about G8KEPR security and trust
We're transparent about our security practices and compliance status.
Contact our security team for detailed discussions about your requirements.
Run on your own infrastructure • Full data control • Enterprise security