Poison Claude Sells Discounted Access to Anthropic’s Claude LLMs While Capturing Every Customer Prompt
What Happened — Researchers monitoring underground cybercrime forums identified at least six advertisements for “Poison Claude,” a service that sells illicit, discounted access to Anthropic’s Claude large‑language‑model family (Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 4.6). The operator of the service reportedly logs every prompt submitted by paying customers, creating a covert data‑exfiltration channel for proprietary or sensitive queries.
Why It Matters for Compliance & Audit Readiness
- Unauthorized model access bypasses the Access Control requirements of SOC 2 CC6.1 (Logical Access) and creates a hidden data‑exfiltration path that must be documented as a control exception.
- The ability of a third‑party to view every prompt violates privacy‑by‑design expectations under GDPR/CCPA and SOC 2 CC5.2 (Privacy). Continuous monitoring of third‑party API usage is essential to provide audit‑ready evidence that only authorized entities can process data.
- Demonstrating that you have a vetted, monitored AI‑service supply chain (e.g., through Verisq’s CookiePLUS privacy controls) helps close the gap between “use of AI” and “compliance with data‑privacy obligations.”
Who Is Affected — SaaS platforms, fintech applications, health‑tech tools, and any organization that integrates Anthropic’s Claude APIs for customer‑facing or internal workflows.
Recommended Actions
- Inventory all AI model integrations and map them to SOC 2 access‑control and privacy criteria.
- Validate API keys against Anthropic’s official credential issuance process; rotate any keys that may have been exposed.
- Enable logging and real‑time monitoring of model‑request traffic to detect anomalous usage patterns indicative of illicit resale.
- Update privacy notices to reflect the risk of third‑party prompt capture and ensure DSAR readiness.
Source: The Hacker News
Technical Notes
- Attack vector: Stolen or illicitly shared API credentials sold on underground forums.
- Data types at risk: Business logic, proprietary algorithms, personally identifiable information (PII) embedded in prompts, and any confidential business queries.
- No specific CVE; the threat is a supply‑chain abuse of Anthropic’s public API.
Source: same as above