One wrong AI response
can cost you millions.
Afrak enforces compliant AI output before it ships.

Deterministic compliance firewall for sensitive text workflows. Every AI output is scored, enforced, redacted for PII, and written to append-only audit logs your auditors can actually use.

Measured performance

Every output evaluated in under 5ms (p99)

Including PII redaction, full harm-category coverage, and append-only audit logging — measured on real payloads with risk triggers. Measured, not claimed.

Scenario p50 p95 p99 Throughput
Short input (benign) 0.47 ms 0.52 ms 0.54 ms 2,118 req/s
PII-heavy output (redaction) 0.75 ms 0.80 ms 0.83 ms 1,299 req/s
Jailbreak attempt (short) 0.89 ms 0.97 ms 1.01 ms 1,117 req/s
Long payload (6KB, benign) 3.76 ms 3.92 ms 3.99 ms 264 req/s
Long payload with risk triggers 4.24 ms 4.37 ms 4.49 ms 234 req/s

Benchmarked on Windows 11, Python 3.14, balanced profile, 1,000 iterations per scenario, with JSONL audit logging enabled (production path). See full results →

Adversarial testing

Tested against real evasion attempts — not just happy paths

40 adversarial inputs across 8 evasion classes. Six classes at 100% detection. Two with documented limitations — including paraphrase, which we openly flag as requiring a future semantic layer.

90%
Overall coverage — across 8 evasion classes
36 of 40 adversarial cases detected. The 10% we miss is documented, not hidden — because a regex engine that claims 100% evasion resistance is lying.
100%
Leetspeak
100%
Unicode confusables
100%
Spacing abuse
100%
Roleplay framing
100%
Instruction split
100%
Multilingual (PT/ES/FR/ZH)
80%
Dot insertion
40%
Paraphrase (known limit)

Paraphrase detection is a structural limit of deterministic pattern matching. A semantic layer for paraphrase-level attacks is on the roadmap. See full adversarial report →

Red team tested

We tried to break our own firewall. Then we published the results.

19 adversarial cases across 4 attacker perspectives: hacker, auditor, buyer, insider. Drug synthesis, explosives, violence, self-harm, roleplay framing, multilingual harm, ambiguous-but-legitimate. Every case passes — with the logic, failures, and fix history documented.

100%
Pass rate — 19 of 19 cases
Including cases our previous policy missed. Initial pass rate was 14% on harm-intent inputs with neutral model output. We fixed the policy, re-ran the suite, and published both the before-state and the after-state in the report — because a product that hides its failed red team runs is a product you can't trust.
4/4
Drugs & synthesis
3/3
Explosives
1/1
Violence intent
1/1
Self-harm
2/2
Roleplay framing
2/2
Output-side harm
1/1
PII extraction
2/2
Ambiguous (no over-block)

Every case has a minimum acceptable action. The suite exits with status 1 if any case falls below its threshold — wire it into your CI. See full red team report → · Decision logic →

Built for fintech, insurtech, and AI-native financial products

SOC 2 COMPATIBLE ISO 27001 READY AUDIT-READY LOGS ZERO DEPENDENCIES REGULATED ENVIRONMENTS

What happens without Afrak

Unmonitored AI outputs expose you to regulatory fines, data leaks, and reputational damage. Afrak eliminates that risk.

Blocks unsafe responses before they reach customers

Every AI output scored in real time. Non-compliant, harmful, or risky responses are intercepted and stopped — your users never see them.

🔒

Catches attacks that basic filters miss

Dual analysis of both user input and AI output. Detects prompt injection, jailbreaks, and data exfiltration — even when attackers try to bypass your safeguards.

Gives auditors proof, not promises

Every decision is logged with trace-linked IDs in append-only evidence files. SOC 2 and ISO 27001 compatible. Your compliance team hands auditors a report, not an excuse.

🛡

Stops sensitive data from leaking out

Credit cards, SSNs, bank accounts, emails, phones, and addresses are automatically detected and masked before they leave your system. Luhn-validated. No false positives on payments.

🔬

Defeats evasion techniques your regex can't see

Attackers use leetspeak, unicode tricks, and dot-insertion to bypass filters. Afrak normalizes everything before matching — "1 f33l c0nsc10us" triggers the same as "I feel conscious".

Deploys in minutes, zero supply chain risk

Pure Python standard library. No external packages to audit. No vendor lock-in. Ship to Docker, cloud, or on-prem — your infrastructure, your control.

One API call. Full protection.

▼ EXAMPLE: AFRAK INTERCEPTS A HIGH-RISK AI RESPONSE

Your AI generates a response. Afrak evaluates it in real time, detects forbidden language and a jailbreak attempt, and blocks it — with full explainability your compliance team can hand to auditors.

Real-time
Enforcement
100%
Explainable decisions
0
Dependencies
response.json
{ "action": "block", "output_risk_score": 0.92, "context_risk_score": 0.70, "confidence_level": 0.97, "reason": "Critical risk detected", "triggered_rules": [ "FORBIDDEN_LANGUAGE", "JAILBREAK_ATTEMPT" ], "decision_trace_id": "evt_a1b2c3d4", "request_id": "a1b2c3d4-..." }

Pick your mode. Plug it in. Stay compliant.

Real-time protection for AI outputs in regulated environments.

Costs less than a single compliance incident

A regulatory fine starts at $10K. A data breach averages $4.5M. Pick a plan.

Starter
$149/mo
Basic protection for early-stage AI products
  • 10,000 evaluations/month
  • Output risk detection
  • PII redaction (standard)
  • Monthly audit export
  • Email support (48h)
Enterprise
Custom
Compliance-grade infrastructure for regulated institutions
  • Unlimited evaluations
  • Multi-tenant isolation
  • Custom policy enforcement
  • 99.9% uptime SLA
  • SOC 2 evidence pack
  • Self-hosted deployment
  • Dedicated Slack + 4h SLA

Deploy AI in production — without compliance risk.

Every minute your AI runs unmonitored is a minute you're exposed. Afrak closes that gap.