Poison Claude Sells Discounted Claude Access While Its Operator Sees Every Customer Prompt

Summary

Illicit services offering discounted AI model access are emerging on cybercrime forums. Poison Claude is one such service, exploiting free credits to provide cheaper access to Anthropic's LLMs. These services pose privacy risks as they can view all user prompts and may lead to data leaks or the provision of inferior models.

Key Takeaways

1. Cybersecurity researchers have discovered more than half-a-dozen services advertisements for illegal access to artificial intelligence (AI) models on underground cybercrime forums and messaging platforms.
2. One such service, Poison Claude, claims to offer access to Anthropic's large language models (LLMs), including Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 4.6.
3. Advertisements for Poison Claude explain how the service can offer the cheap tokens: by taking advantage of free bonus credits, such as the US$100 bonus credit on AWS for Bedrock accounts.

Statement Breakdown

• Claimed Facts: 60% of statements the article presents as facts
• Opinions: 20% of statements classified as editorial or subjective
• Claims: 20% of statements surfaced for additional reader evaluation

Credibility & Bias Reasoning

Credibility assessment: The article relies on research from a cybersecurity company and quotes their findings. While it presents factual information about illicit services, it also includes speculative elements regarding motivations and potential future abuses. The information is presented clearly, but the topic itself involves illicit activities.

Bias assessment: Informative Cybersecurity Focus. The article's primary focus is on reporting cybersecurity threats and research findings. It maintains an objective tone, detailing the mechanics and risks of illegal AI access services without advocating for a particular viewpoint or political stance.

Note: This article details illicit activities and cybersecurity research. While factual, it touches upon potentially sensitive and evolving threats in the AI landscape.

Credibility flag: Investigative Cybersecurity Report

Claimed Facts (7)

• This is a direct statement of fact presented by the article's reporting.
• This states a specific claim made by the service 'Poison Claude'.
• This is a factual explanation of how the service operates, attributed to researchers.
• This is a direct quote from the service's website, presented as factual information about its operations.
• This is a factual statement about the payment methods used by the service.
• This is a factual report of a security vulnerability and its findings, attributed to a security company.
• This presents information about another similar service, including user estimates and claimed offerings.

Opinions (5)

• This statement presents a reasoned perspective on user motivations and inherent risks, which is an analytical opinion.
• This statement outlines potential negative outcomes, which are speculative possibilities rather than confirmed events.
• While attributed to Okta, this statement explains a general operational characteristic and its implication, which can be seen as an analytical opinion on privacy concerns.
• This is a direct statement of concern regarding privacy, which is an opinion on the implications of the described service.
• This statement provides context and an interpretation of market trends, which is an analytical opinion.

Claims (6)

• While the use of CDN is factual, the claim that it's *solely* to conceal the IP address is an assumption about intent, and the domain itself is presented in a way that suggests it's part of the illicit activity.
• The claim that Cloudflare 'declined to take action' is an interpretation of Cloudflare's response and could be subjective or incomplete without direct confirmation from Cloudflare.
• This is a claim of accusation, which is a serious allegation and presented without direct evidence of the campaigns' success or scale, relying on Anthropic's statement.
• This claim is based on a Reuters report, which itself may have varying degrees of substantiation. The direct link and intent of 'advancing defense capabilities' can be a strong interpretation.
• The phrase 'evidence shows' is vague, and the connection between abusing free AI trials and synthetic identity creation at scale, while plausible, is presented as a direct causal link without detailed evidence within the article.
• While attributed to Okta, the statement about 'rising bot activity' is a broad claim that could be difficult to definitively prove and quantify without specific data presented.

Key Sources

• The Hacker News — Media
• Jeremy Kirk — Researcher, Okta
• Mathew Woodyard — Researcher, Okta
• Okta — Identity Security Company
• Poison Claude — Illegal AI Access Service
• Anthropic — AI Company
• Reuters — News Agency