AI pentesting for your company
AI penetration testing is here: AI agents find vulnerabilities on their own and prove them all the way to the exploit. In the independent BB Arena benchmark, the top model fully solves 62 of 100 realistic web and API labs, without source code or hints. With 2342.ai you test your own systems with the same capabilities: through one API, with budget, roles and documented data paths.
The underestimated risk
Internal testing finds exactly what needs protecting: HR files, customer records, emails, credentials. When an employee pastes those findings into a personal ChatGPT or Claude subscription for analysis, personal data leaves the company: without a data processing agreement, without control over retention and, depending on plan and settings, with possible use for training. At the same time, the newest frontier models increasingly stop security tasks. In BB Arena, provider safeguards stopped 9 of 25 configurations, in up to 97 of 100 labs. Reaching for some uncensored API without a contract then makes the problem bigger.
Findings with personal data
Analyze logs, responses and records containing personal data with models the catalog lists with an EU data path. According to BB Arena, several of them handle security tasks without a single refusal.
Security models
For exploit development, payloads and attack chains against your own systems, specialized security LLMs and abliterated models are available, such as CyberGLM, CyberKimi and Abliterated Model Large. The inference provider stores no data (zero data retention). Processing takes place in Norway, Iceland and the United States, so these models are not intended for personal data.
Control
Agents such as Claude Code, OpenCode or your own test harness connect through OpenAI- or Anthropic-compatible endpoints. Every test runs with the organization's own API keys; spending limit and usage stay visible.
How to start
Only test your own systems or systems with written authorization. For hosted services, the provider's testing rules apply as well.
The organization admin enables external models for the organization. The security models are then available with the organization's API keys.
Attacks with security models, analysis of personal data findings with EU data path models, reporting and remediation in the workspace.
No. The models are intended for your own systems and authorized engagements. Attacking systems without the owner's permission can be a criminal offense. The terms govern permitted use in the section "Security and abliterated models".
Abliteration identifies the direction in a model's activations that triggers refusals and removes it from the weights. The model no longer refuses tasks, while its other capabilities stay largely intact. For security testing, that means exploit code, payloads and attack chains without interruptions, but also without built-in guardrails. That is why we offer them only through the API and only to organizations that have enabled them.
Personal subscriptions are consumer contracts: no data processing agreement with your company, no control over retention and, depending on settings, use for training. The newest models also frequently stop security tasks. In BB Arena, Claude Opus 5.5 stopped in 51 of 100 labs, ChatGPT 6.1 Sol in 67.
Technical content without personal data: code, configurations, requests, payloads. For findings with personal data, use models with an EU data path, for example Claude Opus 5, GLM 5.3 and ChatGPT 5.6 Sol.
Not yet. Our benchmark page only shows published results, and none are available for the security models in our catalog so far.
No. They are only available through the API, using the keys of an organization whose admin has enabled external models.