PentestLLM is a model gateway designed for authorised penetration testing, red-team engagements, bug-bounty research and capture-the-flag practice. It gives security practitioners one prepaid key for security-tuned, abliterated and uncensored models through OpenAI- and Anthropic-compatible interfaces. The service is operated by Uplink Switch LLC and repeatedly limits its intended use to systems the customer owns or has written permission to test.

Security workflows

PentestLLM works alongside established tools rather than replacing them. It can read and rank output from scanners such as Nuclei, review notes or code, help draft a proof of concept for a controlled lab and turn verified findings into report language. The tester remains responsible for scope, validation and every action executed against a target.

  • A large catalogue of security-tuned, abliterated and other open-weight models
  • OpenAI-compatible API, Anthropic Messages support and coding-agent guides
  • Automatic routing across security tiers and configurable fallback models
  • Web search, web fetch, remote MCP and code-sandbox tools within a request
  • PDF and document input with redaction options for secrets and personal data
  • Session traces, usage export and observability integrations
  • Embeddings, audio, reranking and batch processing on the same key
  • Shield decoys intended to trap and alert on attacking automated agents

Models and pricing

The live catalogue listed more than 370 models at review time, with model health checked daily. The company distinguishes security-tuned builds from models whose refusal behaviour has been reduced. Catalogue size is not a measure of pentesting quality, and the service does not claim that every model is trained specifically for security.

Pay-as-you-go access starts with a US$10 prepaid top-up; balances do not expire. Text models are billed by separate input and output prices per million tokens. On 21 November 2026 the lowest advertised output rate was US$0.08 per million tokens, while larger and specialist models cost substantially more. One-time US$200, US$500 and US$1,000 top-ups advertised 5, 10 and 15 per cent bonus credit respectively.

Recurring usage plans are Plus at US$10 for US$20 of monthly usage, Pro at US$25 for US$50, Max at US$100 for US$200 and Ultra at US$250 for US$500. They renew by card until switched off. Failed requests cost nothing, but web search and other tools may carry separate charges. The live price index should be checked for the selected model before a large engagement.

Scope, privacy and risk

PentestLLM says request and response bodies are not retained by default, although seven-day replay can be enabled per key. It stores usage metadata, token counts and charges, while third-party model hosts process the content needed to answer. New messages are also classified by OpenAI's moderation API. Security teams handling client data should reconcile those processors with their engagement terms before uploading source code, scan output or unpublished vulnerabilities.

The acceptable-use rules prohibit testing without authorisation, malware deployment and other harmful activity. Reduced model refusals do not override the law, a rules-of-engagement document or a bug-bounty programme's scope.

Our assessment

PentestLLM can be useful when a legitimate security team needs model choice and fewer false refusals around dual-use terminology. Its strongest fit is analysis, prioritisation, lab-based drafting and reporting, where an experienced tester can verify every claim. It should not be treated as an autonomous penetration tester or evidence that an unverified finding is exploitable.

Prospective users should compare models on their own authorised examples, inspect how prompts are routed, restrict keys and spending, and keep human approval before any active step. The platform's broad catalogue and low entry price aid evaluation, but buyers still need their own governance, client consent and evidence-handling controls.