Developer-controlled security AI

LLMs for security teams and legitimate penetration testing.

Authorized security research needs full model capability. Governance still needs control.

OpenAI-compatible API + Policy Gateway lets you run legitimate penetration testing, red team, and security research workflows that often trigger generic “hacking” filters, while enforcing your own rules: redact, rewrite, escalate, audit.

OpenAI-compatibleNo prompt / output retention by defaultPolicy-as-code controlsAudit logs exportable to SIEM~$3 / 1M tokens

Prompts and outputs are processed transiently and not stored by default. Payloads are never used for model training.

Security evaluation

See whether Abliteration fits your red-team workflow before you commit engineering time.

Share the workflow you want to unblock. We will map it to an API path, retention posture, and Policy Gateway setup that can survive internal review.

Authorized testing scope

Policy and audit design

API rollout plan

Request a security workflow review

Best for AppSec, red-team, AI safety, and security research teams evaluating governed LLM access.

Security eval review

Share the authorized security workflow you want to run without provider-side refusals, plus any governance or audit requirements.

By submitting you agree to our Privacy Policy.

The problem

Why standard LLMs fail red teams.

  • Generic "hacking" filters block lawful penetration testing prompts that security teams are authorized to run.
  • Refusals derail automated pipelines and produce inconsistent results.
  • Provider moderation changes create regressions you can't control.
  • Lack of auditability creates client-trust issues.
  • Data retention and training concerns block procurement.
  • Multi-tenant controls are hard to implement yourself.

Abliteration gives you full capability and an enforcement layer you control.

Capabilities

How Abliteration helps.

Outcome-driven capabilities mapped to your security workflow needs.

Fewer refusals, more reliable automation

Abliteration is a model-editing approach designed to reduce refusal behavior — more stable than prompt jailbreaking across variations. 3/100 refusal rate on harmful-behaviors eval.

Policy Gateway = your rules, enforced in real time

Policy-as-code with millisecond enforcement. Predictable outcomes: rewrite, redact, escalate, or refuse. Shadow mode, canary rollouts, and auto-rollback.

Audit trails your enterprise clients will accept

Structured decision metadata + SIEM exports to Splunk HEC, Datadog Logs, Elastic, S3, and Azure Monitor. Includes policy_id, user/project tags, reason codes, triggered categories.

Client isolation + quotas by default

Create a project per client. Issue scoped keys. Enforce per-user and per-project quotas. Revoke independently. Every engagement gets its own risk profile.

Privacy-first handling for sensitive engagements

No prompt/output retention by default. Never used for training. Internal network details, vulnerability evidence, and client artifacts stay yours.

Drop-in integration

OpenAI-compatible /v1/chat/completions. Change base URL + key, keep your SDK and request schema. Supports streaming, structured JSON output, and function calling.

Use cases

Where this fits in your security product.

Finding triage at scale

Turn large volumes of scanner outputs into structured summaries, deduped findings, and prioritized remediation narratives.

Report automation

Generate consistent executive summaries, risk narratives, and remediation guidance with a standard format across analysts.

Playbook generation

Create reusable test checklists and verification guidance for security controls — defensive and validation — without leaking sensitive client specifics.

Agentic workflows with governance

Build agents that call internal tools with allowlists and audits for tool calls. Policy Gateway enforces rules on every tool invocation through the MCP guard.

Policy example

Security research policy example.

See the governance knobs. Project isolation, quotas, audit logging, redaction, and escalation — not blunt refusal.

policy.json
{
  "policy_id": "secops-redteam-v2",
  "project_id": "client-acme-2026",
  "rules": {
    "categories": {
      "pii_leak": "redact",
      "credential_exposure": "redact",
      "scope_violation": "escalate",
      "exploit_generation": "allow",
      "vulnerability_analysis": "allow",
      "social_engineering_templates": "escalate"
    },
    "allowlist": [
      "penetration testing",
      "vulnerability assessment",
      "security audit",
      "red team exercise"
    ],
    "denylist": [
      "real target names outside scope"
    ]
  },
  "quotas": {
    "per_user_daily_tokens": 2000000,
    "per_project_monthly_tokens": 50000000
  },
  "audit": {
    "enabled": true,
    "export": "splunk_hec",
    "include_decision_metadata": true
  },
  "rollout": {
    "mode": "enforce",
    "shadow_percent": 0,
    "canary_percent": 100
  }
}
Integration

30 seconds to value.

Works with your current SDK. Change the base URL and API key — everything else stays the same.

Python — base URL swap
import openai, os

client = openai.OpenAI(
    base_url="https://api.abliteration.ai/v1",
    api_key=os.environ["ABLIT_KEY"],
)

response = client.chat.completions.create(
    model="abliterated-model",
    messages=[{"role": "user", "content": "Analyze these scan results..."}],
    stream=True,
)

for chunk in response:
    print(chunk.choices[0].delta.content or "", end="")
StreamingStructured JSON outputFunction / tool calling

Trust, boundaries, and acceptable use

  • For authorized security testing and research only. You must have explicit written authorization for any target systems.
  • You must comply with applicable laws and scope-of-work agreements.
  • We investigate abuse and can suspend accounts that violate our terms of service.

Abliteration provides capability + governance. Policy-as-code controls, scoped keys, quotas, audit logs, and SIEM exports give you and your clients the confidence to deploy responsibly.

Ready to run the model without the filter friction?

Grab a key, swap the base URL, and plug Policy Gateway in when you're ready for audits and rollouts.

FAQ

Frequently asked questions.

Can I use abliteration.ai for automated penetration testing?

Yes. abliteration.ai provides developer-controlled models with reduced refusal behavior, designed for authorized security testing workflows. Policy Gateway adds governance controls so you can enforce your own rules rather than relying on provider defaults.

What if legitimate penetration testing prompts trigger hacking filters on other APIs?

That is a common problem. abliteration.ai is designed for authorized security teams whose lawful pentesting and red-team prompts trigger generic hacking filters on mainstream APIs. Policy Gateway lets your organization own the rules while keeping auditability and scoped access controls.

Will my security research prompts be stored or used for training?

No. Prompts and outputs are processed transiently and not stored by default. Your data is never used for model training. Only operational telemetry (token counts, timestamps, error codes) is retained for billing.

How does Policy Gateway help with red team governance?

Policy Gateway provides policy-as-code enforcement with rewrite, redact, escalate, or refuse outcomes. You can isolate client engagements with per-project keys and quotas, export audit logs to your SIEM, and test policy changes with shadow mode and canary rollouts.

How do I integrate abliteration.ai into my security tooling?

abliteration.ai is OpenAI-compatible. Change your base URL to https://api.abliteration.ai/v1 and swap your API key. Your existing SDK, request schema, and streaming code work without changes.

Can I prove AI governance to my enterprise clients?

Yes. Policy Gateway logs every decision with structured metadata including policy_id, user/project tags, reason codes, and triggered categories. Export audit logs to Splunk HEC, Datadog Logs, Elastic, Amazon S3, or Azure Monitor.

How does multi-tenant isolation work for client engagements?

Create a project per client engagement, issue scoped API keys, enforce per-user and per-project quotas, and revoke access independently. Each engagement gets its own audit trail and budget controls.