Skip to content
AI Security SentinelAI Security Sentinel

4. Abliteration AI — Guardrail-Removed Models

Contents5 numbered sections

4.1 Identity

Section titled “4.1 Identity”

4.2 Models & Functionalities

Section titled “4.2 Models & Functionalities”
Model ID Base Ctx / Max out Modalities Notes
abliterated-model undisclosed general 262K / 262K Text+image in, video on Chat only Default multimodal
abliterated-model-large GLM-5.2 abliterated + fine-tuned adversarial 1M / 1M Text-only Previous large
abliterated-model-large-v2 GLM-5.3 abliterated, FP8 1M / 1M Text-only, 3 reasoning modes low/high/max Current default large

Sources: https://docs.abliteration.ai/models.md, https://abliteration.ai/blog/introducing-abliterated-model-large-v2.

Vendor benchmarks (not independently verified):

  • base: mmlu_pro 82.1, gpqa 73.1, aime_2025 83.7, mmmu_pro 68.1, refusals 3/100 harmful_behaviors.
  • large (GLM-5.2): SWE-bench Verified 81.2%, Terminal-Bench 2.1 80.1% (base 81.0%), AgentHarm 86.2% zero refusals (“highest published”), AgentDojo 97.5% benign / 34.29% under injection / 57.86% targeted ASR, CyberGym 84.2% vs GPT-5.5 81.8%.
  • large-v2 (GLM-5.3): CyberGym 84.5% SOTA claim, Terminal-Bench 4.0 41.8%, ExploitGym 29→105 tasks in 2h, ExploitBench 24.4→54.4 (2×), #3 Terminal-Bench 4.0 behind Opus 5 and Fable.

Platform layers (/platform): (1) Hosted abliterated LLM — no per-request refusals on cybersecurity/red-team/training-data/agent workloads, (2) Policy Gateway — policy-as-code allow/refuse/rewrite/redact/escalate + reason code to SIEM, (3) Training-data console — batch dataset gen, paid 3-row preview, export HF/Kaggle/S3/GCS/Azure.

Marketed offensive uses: pen-test, CVE repro, exploit dev, malware analysis, phishing pretexts, password-stealer code, SQLi demo in console; AI red-teaming (jailbreaks, direct+indirect prompt injection via PDF/email RAG exfiltration, tool-misuse coercion, model-stealing probes, DAN); Trust & Safety synthetic data (coded harassment, 20 harassment examples, 50 phishing emails); ML research / defense-government.

Caps matrix: Chat/Responses/Messages ✓, streaming ✓, tool calling ✓, structured JSON ✓, reasoning-effort ✓, disable/hide reasoning ✓, reasoning trace ✓, web search ✓ ($8/1k), web fetch Anthropic-only ✓, token counting ✓, safety filtering via flagged_categories ✓, auto prompt caching ✓. No self-host — hosted-only. Self-host alternative is community weights + vLLM (e.g. dealignai/GLM-5.3-ABLITERATED-NVFP4, Securelayer7/Qwen3.8-27B-Uncensored-Abliterated, huihui-ai/BaronLLM_Offensive_Security-abliterated-GGUF).

4.3 Pricing

Section titled “4.3 Pricing”

Source of truth https://abliteration.ai/pricing + https://docs.abliteration.ai/pricing (pricing page overrides older $1 copy).

  • Subs (monthly reset, no rollover; prepaid stacks, never expires): Developer $20/mo 2.5% discount (solo, pay-as-you-go, project-limited keys, auto-reload); Growth $50/mo 5% (higher limits, spend controls, audit logs, team, email); Scale $200/mo 10% ($200 included credit, highest limits, priority); Enterprise Custom (dedicated throughput, custom routing/region, volume, contracts/compliance, EKM, Policy Gateway). Free preview 1 credit ≈500 full-price base tokens, no card.
  • Per-token USD per 1M (input+cached+output billed separately; image/video as token equiv on base only):
Model Input Cached Output
abliterated-model 256K $3.00 $0.30 $3.00
abliterated-model-large 1M $5.00 $0.50 $5.00
abliterated-model-large-v2 1M $5.00 $0.30-0.50* $5.00

*Docs $0.50 vs blog $0.30 — confirm at checkout. Effective ~$3/1M total tokens ($10≈3.3M). Web search $8/1k + tokens. Old $1-input copy superseded.

  • Listings: Slashdot confirms $20 start, free trial yes, 0 ratings.

4.4 API

Section titled “4.4 API”

Hosted-only; EU residency option in-region (zero-retention). Quickstart:

Terminal window
export ABLIT_KEY=ak_YOUR_API_KEY
curl https://api.abliteration.ai/v1/chat/completions \
  -H "Authorization: Bearer $ABLIT_KEY" -H "Content-Type: application/json" \
  -d '{"model":"abliterated-model-large-v2","messages":[{"role":"user","content":"Hello"}]}'

Python OpenAI(base_url="https://api.abliteration.ai/v1", api_key=os.environ["ABLIT_KEY"]), Node same, key prefix ak_.... Must include /v1, exact model, messages; 401 bad key, 404 missing /v1, 400 bad shape, 429 backoff.

Endpoints: POST /v1/chat/completions, /v1/responses, /v1/messages (Anthropic), /v1/messages/count_tokens, GET /v1/models, GET /credits/balance, /policy/*. Multimodal content:[{type:text},{type:image_url},{type:video_url}], HTTPS or data URLs, SSRF-guarded. Streaming SSE. OpenAPI https://api.abliteration.ai/openapi.json.

Integrations: LangChain (ChatOpenAI baseURL swap), LlamaIndex (OpenAILike), Vercel AI SDK (@abliterationai/ai-sdk-provider), Cloudflare Workers, Claude Code (2 env vars), Codex CLI, Pi agent, CC Switch, CLIProxyAPI, OpenCode built-in provider, CyberStrike fork, Strix pentest, Hermes Agent, OpenClaw plugin, Promptfoo built-in provider for eval/red-team, Giskard scanner backend, Mastra router (ABLIT_KEY). Examples abliterationai/abliteration-examples (6★). Policy Gateway POST /policy/chat/completions with policy_id, policy_user, project ID; connectors Splunk HEC, Datadog, Elastic/OpenSearch, Azure Monitor, S3/Blob/GCS/B2/R2, webhook, OTel. Rate limits tier = max(sub plan, lifetime-spend tier).

4.5 Legal / Ethical

Section titled “4.5 Legal / Ethical”
  • Terms (2026-01-08): 18+, accurate registration, safeguard ak_ keys, comply laws/export/third-party terms, credits non-refundable/non-transferable, Acceptable Use: “No illegal, harmful, or high-risk activities (including anything that may harm people or infrastructure). No evasion/probing/disruption. May investigate/suspend.” Payload transient, no default storage; telemetry (counts/timestamps/endpoints/codes) for billing. As-is, no warranty, liability capped greater of $100 or prior 3mo fees, Delaware venue. Termination via help@.
  • Data (2026-05-29): Zero retention by default (vs ~30d elsewhere); prompts/completions/images in-memory then discarded, never for training, redacted from error logs; retained token counts/timestamps/status/model/billing (lifetime + hold). Exceptions: web search/fetch sends query/URL to third-party (their retention, zero-retention void if enabled); training-data generator stores datasets you generate (deletable, never trains vendor); Policy Gateway stores rules + content-free metadata only. EU residency in-region, EKM enterprise.
  • Authorized-testing framing (/security-testing): “For authorized security testing and research only. Explicit written authorization required. Comply with laws/scope. We investigate/suspend.” Per-client projects, scoped keys, quotas, revocation, SIEM audit, shadow/canary/rollback, PII redact/rewrite/escalate. Marketing: “Unrestricted. Not ungoverned”, “frontier intelligence, your guardrails”.
  • Gap to flag: ToS bans illegal/harmful/high-risk yet homepage sells offensive cyber/exploit/malware analysis and LessWrong/TechCrunch show pathogen + stealer compliance with only card/email check, no use-case screening, no monitoring, filters only self-harm/CSAM. Authorization burden + Policy Gateway (Enterprise-only) pushed to you. No public DPA/BAA; deletion via email only.