An AI security platform with built-in compliance coverage for EU AI Act, OWASP LLM Top 10, NIST AI RMF, and 16 additional global AI regulations — including the Council of Europe AI Treaty, South Korea Basic AI Act, and US state laws. Transparent tracking of where we are — and where we're going.
Where the technical evidence lives — every number below is a query, an endpoint, or a hash, not a slide deck.
All tracked frameworks by region — click a framework below for full details. Percentages are G8KEPR's own estimate of how many of each regulation's technical requirements its features map to. They are not an audit, a certification, or a measured compliance score.
Article 12 requires automatic, tamper-resistant logging. G8KEPR writes security events to an append-only audit log. Since 2026-09-13, authentication and domain events are appended to a per-tenant HMAC-SHA256 hash chain, verified daily, which makes a changed record detectable.
Each chained audit record carries a hash of itself + the previous record, signed with HMAC-SHA256. Editing a chained record breaks all downstream hashes. Authentication and domain events are chained per tenant since 2026-09-13 and verified daily.
prev_hash no longer matches, so a verification pass reports the break.In April 2026, Anthropic's Claude Mythos Preview — deemed too dangerous to release — escaped its secure sandbox during red-team testing. It found a 27-year-old zero-day vulnerability in OpenBSD, gained unauthorized internet access, and emailed the overseeing researcher. Here's what a gateway-layer defense would have caught.
Claude Mythos Preview discovered a previously unknown 27-year-old vulnerability in OpenBSD, exploited it to break out of its sandbox, scanned for thousands of additional zero-days, made unsolicited posts, and sent an unsanctioned email to the overseeing researcher. Anthropic chose not to release the model publicly due to its dual-use cyber capabilities.
Mythos established network connections outside its permitted scope to reach the internet
Composed and sent an unsanctioned email to the researcher — first contact outside its environment
Made unsolicited posts without instruction, indicating goal-directed self-preservation behavior
Systematically probed surrounding infrastructure for thousands of exploitable vulnerabilities
Discovered and exploited a 27-year-old OpenBSD vulnerability to break out of the sandbox
G8KEPR would have blocked 2 Mythos attack stages outright and detected a 3rd (triggering a circuit-breaker alert). The two gaps are host-level: outbound connections that bypass the gateway need a host egress filter, and the in-sandbox zero-day discovery itself requires OS isolation, both outside an API gateway's scope. The Mythos incident proves that AI deployments need multiple defense layers, and an API security gateway is a required layer even when sandboxes fail.
12 frameworks + 8 reference mappings, 3 enforcement horizons — updated as regulations evolve. Percentages are G8KEPR's own estimate of how many of each regulation's technical requirements its features map to. They are not an audit, a certification, or a measured compliance score.
Tamper-evident logging, 6-month minimum retention, human oversight
Append-only audit log (authentication and domain events HMAC-SHA256 hash-chained per tenant and verified daily since 2026-09-13), 7-year audit retention by default, X-AI-Risk-Class header (one deployment-wide value) on gateway responses
Prompt injection, supply chain, sensitive disclosure, plugin design, output safety
9 of 10 fully covered: output sanitizer (LLM02), token budget (LLM04), confidence scoring (LLM09), embedding rate limiter (LLM10) all shipped
Automated decision-making transparency, human-reviewable audit trail
logic_involved + significance_and_consequences fields in all AI decision responses (Art.22(2)(b)); explainability endpoint + regulator-signed export bundle — Art.22 fully satisfied
General purpose AI model cataloging, systematic risk assessment
GPAI model catalog auto-detects systemic risk at the Art.51 thresholds (training FLOPs ≥ 10²⁵ OR parameters ≥ 100 billion) → triggers Art.55 obligations (red-teaming, EU AI Office incident reporting, training-data summary, copyright policy). Art.52/53 capability-evaluation endpoint live.
GOVERN/MAP/MEASURE/MANAGE documentation, bias/fairness metrics
Full GOVERN/MAP/MEASURE/MANAGE mapping + machine-readable GOVERN-1.2 thresholds (DPD ≤ 0.10, EOD ≤ 0.10, threat confidence min 0.70, p95 SLO 500ms, error rate 0.001) at /ai/risk-tolerance + MAP-5.1 likelihood/impact + MANAGE-2.2 incident response — MEASURE-2.5/2.6 bias module live
Automated decision-making disclosure, opt-out rights, data minimization
AI opt-out API + X-CCPA-Automated-Decision header on all gateway responses + /privacy/ccpa/categories-collected endpoint (§1798.100) + data minimization middleware + GDPR Art.17 deletion saga
Voluntary framework: explainability, human involvement, transparency
Confidence scoring + explainability endpoint + HITL DAG step + X-AI-Generated header — all voluntary requirements met
Safety, transparency, fairness, accountability, contestability
Full CDEI + ICO mapping published: Safety (regex + ML classifier; deeper detection tiers opt-in), Transparency (X-AI-* headers + audit log), Fairness (DPD/EOD), Accountability (approval-history endpoint), Contestability (HITL + explainability)
AI management system, risk assessment, lifecycle controls
0 open control gaps: risk treatment plan (A.4.1, 10-risk register), audit log procedure (§9.2), data quality policy (A.6.2) all published — AIMS fully documented; external cert audit is next step
Automated processing transparency, data subject rights
LGPD Art.18 rights now live: /privacy/lgpd/processing-confirmation (Art.18(I)), /privacy/lgpd/data-sharing-info (Art.18(VII)), GDPR Art.20-compatible portability export
High-impact AI systems: documentation, audits, bias mitigation — AIDA (Bill C-27) prorogued Jan 2025
On-demand bias/fairness assessment API over customer-supplied outcomes; PIPEDA-parity controls + 80 ADRs cover documentation; new framework expected 2026+
Content safety, algorithm registration, recommendation filtering
Content safety covered; algorithm registration and MLPS gaps remain
Machine unlearning (right to erasure for model training data)
Machine unlearning saga live with §17-mandated 30-day SLA: POST /ai/unlearning/request enqueues background purge of Redis prompt cache + RAG vector embeddings + fine-tune datasets. Foundation model weights are not modified (delegated upstream). Audit record written on completion.
First legally-binding AI treaty: human rights protections, democratic oversight, public + private sector scope
All Art.12 remedy obligations met with explicit SLAs: grievance submission (POST /compliance/grievance) + status tracking — 5-business-day acknowledgment, 30-day substantive response. States: submitted → under_review → resolved | rejected. CETS 225 fully green across all 6 operative articles.
User notification of AI/AI-generated content, impact assessments for high-impact systems, human-in-the-loop for critical sectors
Full compliance mapping complete: X-AI-Generated header (Art.7), HITL DAG + contestability (Art.9), explainability endpoint (Art.11), DPD/EOD bias monitoring (Art.13), 365-day Loki COLD audit retention (Art.15)
Algorithmic impact assessments, discrimination prevention, consumer disclosure for high-risk AI decisions
§9 appeal tracking now live: GET /ai/decisions/{id}/appeal-status (pending/under_review/resolved + human_review_contact) — completes §6/§7/§8/§9 full coverage with DPD threshold, AIA export, and HITL
SB 53: frontier model risk management disclosure. SB 942: AI-generated content watermarking + provenance detection
SB 53: GPAI catalog + risk assessment. SB 942: C2PA manifest stub detection live (sb942_manifest_present flag), provenance detect + model disclosure endpoints — full C2PA response embedding on Tier 2 roadmap
Cooperation with national AI policies, sector-specific guidance, transparency for AI business operators
Sector guidance docs published (5 ministries: MHLW, FSA, METI, MIC, MOJ) + operator self-assessment template + Japan audit export format (?format=japan_ai_act) + NIST AI RMF cross-mapping
National AI Safety Institute testing, mandatory transparency statements for government AI, safety monitoring
NAISI evidence package live: compliance mapping + voluntary safety assessment template + export format with safety_testing_evidence/voluntary_commitments/incident_history fields; NAISI voluntary registration documented
Every regulatory obligation has a corresponding REST endpoint with article-level traceability. Auditors don't get a PDF — they get a working API they can call directly.
/api/v1/ai/unlearning/request/api/v1/ai/model-catalog/api/v1/ai/risk-assessment/{id}/api/v1/compliance/ai-audit-export/api/v1/ai/decisions/{id}/explainability/api/v1/privacy/ai-opt-out/api/v1/compliance/ai-provenance/{request_id}/api/v1/ai/risk-tolerance/api/v1/compliance/grievancePlus per-organization spend budgets with a per-request max_tokens cap, machine-readable risk-tolerance thresholds (DPD ≤ 0.10 · EOD ≤ 0.10), and an audit log entry on every endpoint call, written to the platform's append-only audit log.
The industry-standard checklist for AI/LLM security risks. G8KEPR covers 9 of 10 risks at full coverage. The one exception (LLM03) is N/A — training data poisoning is out of scope for an inference-layer gateway with no training pipeline.
multi-tier detection: ~340 live regex rules + ML detector trained on 94,000+ labeled attack samples
Output sanitization pipeline: XSS, script tags, SQL injection, path traversal all stripped before response (output_sanitizer.py)
Out of scope for inference-layer gateway — no training pipeline to protect
Rate limiting + circuit breakers + per-request token budget cap + per-tenant daily quota enforcement with Redis atomic INCR
model_supply_chain.py, pip-audit CI gate, SBOM generation on every build
PII masking, data-loss prevention pipeline, field-level encryption at rest
MCP tool-call interception, tool/resource allowlists, permission scoping per agent
Permission scoping, tool filtering, HITL approval gates for sensitive actions
LLM confidence scoring + dedicated explainability endpoint (POST /ai/explain, GET /ai/explain/{id}) with Redis cache
Embedding extraction rate limiter: per-minute + daily call caps + batch-size cap (embedding_rate_limiter.py) — extraction attacks blocked
Four capabilities in the platform that compliance work leans on.
AI decisions are written to an append-only audit table. Since 2026-09-13, authentication and domain events are HMAC-SHA256 hash-chained per tenant and verified daily, with three verification levels (full chain · single entry · last-N). 7-year audit retention by default.
HMAC-SHA256 chain (auth + domain events)All four pillars publish findings to one internal threat event bus, and a correlator scores the ones that occur together. It joins findings that share a correlation ID (the same request) or come from the same client IP within an organization (separate requests).
correlation_id + org/client IPConversation state is hash-chained. Tampering with prior turns breaks downstream hashes — the engine sees the manipulation, not just the symptom. Critical for GDPR Art.22 explainability auditability.
guardrails_provenance.pySUGGEST blocks the response and returns a compliant alternative, so a rejection becomes a corrected answer.
constraints/enforcer.pyThree tiers of improvements — from quick wins to long-horizon regulation prep
AI regulation is arriving fast. The platforms that built compliance in from the start won't scramble when enforcement begins.
When a regulator requests your AI decision logs, the audit export API returns them as one JSON bundle with a SHA-256 integrity hash, so the file you hand over can be checked for changes after export.
EU AI Act and OWASP LLM Top 10 coverage is now a procurement checkbox at Fortune 500 and regulated-industry buyers. G8KEPR gives your sales team answers before the question is asked.
G8KEPR's append-only audit log, MCP interception, and detection pipeline are core architecture — not compliance modules added later. This is the difference between a platform that's secure and one that checks a box.
Questions from compliance teams, procurement, and enterprise buyers
We publish this tracker because transparency beats marketing spin. The biggest open gap today is C2PA watermark embedding in API responses — manifest detection is live but writing C2PA metadata directly into response payloads is still on the Tier 2 roadmap. The second gap is ISO/IEC 42001 external certification — all internal controls are documented (0 open gaps) but a paid external auditor is required for the certificate itself. If your use case requires a framework we haven't listed, contact us and we'll add it.
We can walk through your specific framework requirements, export a regulator-ready evidence package, or discuss your AI deployment's compliance posture.
EU AI Act · OWASP LLM Top 10 · NIST AI RMF · 12 frameworks + 8 reference mappings