Poison Claude - Man-in-the-Middle Reseller Service for Anthropic Claude Access

Technical Analysis

Poison Claude appears to function as a proxy/reseller layer that fronts access to genuine Anthropic Claude API/account sessions, likely obtained via stolen API keys, compromised OAuth tokens, or abused enterprise trial/credit accounts, then resold at a discount on underground forums and messaging platforms. Because all inference requests are routed through the operator's infrastructure, the service can log, inspect, and retain every prompt and response, exposing proprietary data, credentials, internal system prompts, and tool-call payloads submitted by customers. This is functionally a man-in-the-middle (MITM) scheme against LLM traffic rather than a traditional exploit chain, with no CVE associated, but it mirrors risks seen in malicious LLM API proxies and token-theft marketplaces. Organizations using such gray-market access inside automated pipelines, LLM tool-use chains, or RAG/agent frameworks risk leaking API keys, internal documents, retrieved context, and agent tool outputs directly to a hostile third party, and any agent that autonomously calls out to third-party LLM endpoints without verifying provenance is at direct risk of full prompt/response interception and secondary compromise.

Affected Systems

Anthropic Claude API consumers (Opus 4.8, 4.7, 4.6, Sonnet 4.6) accessed through unauthorized reseller/proxy services; organizations or developers using non-official Claude access channels; AI agent frameworks, RAG pipelines, and automation tools configured to call third-party or unofficial Claude endpoints

Indicators of Compromise

• Service name: "Poison Claude"
• Distribution channels: underground cybercrime forums, encrypted messaging platforms (e.g., Telegram-based marketplaces)
• No confirmed hashes, IPs, or domains published at time of reporting

Remediation Steps

1. 1

   Use only official Anthropic API channels

   Ensure all LLM integrations, agents, and automation pipelines authenticate exclusively through Anthropic's official API endpoints and verified enterprise accounts, never through third-party resellers or discounted access brokers.

2. 2

   Audit and rotate API keys

   Review all Claude API keys and credentials in use across agent frameworks and RAG pipelines; rotate any keys that may have been shared, embedded in scripts, or exposed to unofficial services.

3. 3

   Monitor for anomalous account/API usage

   Watch for irregular billing spikes, unfamiliar geographic access patterns, or usage inconsistent with expected agent workloads, which may indicate compromised or resold credentials.

4. 4

   Enforce data handling policies for LLM traffic

   Prohibit sending sensitive, proprietary, or credential-bearing prompts through unauthorized or non-vetted LLM access services within agent tool-use configurations.

5. 5

   Educate teams on gray-market AI access risks

   Brief developers and security teams that discounted or 'unofficial' LLM access services likely involve MITM prompt logging and should never be used for production or agentic workloads.