This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
A Python-based framework called [name withheld]Bot utilized OpenAI's GPT-4o-mini model to generate unique, context-aware spam messages for website contact forms and chat widgets. By tailoring messages to specific websites, the bot successfully bypassed traditional spam filters, targeting over 400,000 domains and successfully spamming at least 80,000 since September 2024. The operation also employed CAPTCHA bypass services and proxy rotation to evade detection, though OpenAI has since disabled the API keys associated with the activity.
SentinelLabs reported that unknown operators used AkiraBot, a Python framework, plus OpenAI's chat API to generate customized SEO spam and bypass CAPTCHAs, posting via SMB websites' contact forms and Reamaze-style chat widgets. Researchers assessed >400k domains were targeted and ≥80k successfully spammed since 09/01/2024. OpenAI said the API key was disabled after disclosure.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
Using AI systems to gain a personal advantage over others such as through cheating, fraud, scams, blackmail or targeted manipulation of beliefs or behavior. Examples include AI-facilitated plagiarism for research or education, impersonating a trusted or fake individual for illegitimate financial benefit, or creating humiliating or sexual imagery.
Also classified as
The framework flooded tens of thousands of website contact forms and chat widgets with unsolicited spam, violating platform policies and community norms.
The bot used AI to generate highly tailored spam messages to deceive specific website owners into purchasing dubious SEO services for the operators' personal gain.
Human
Due to a decision or action made by humans
Intentional
Due to an expected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The incident was caused by human threat actors who intentionally designed and deployed the [name withheld]Bot framework to abuse OpenAI's API for a large-scale spam campaign.
Added categories peak at 2: Minor (Epistemic Harm, Direct)
The [name withheld]Bot incident represents an evolved capability in automated spam operations, leveraging GPT-4o-mini to generate context-aware messages that bypass traditional security filters. While it successfully targeted over 80,000 websites, its national security impact remains negligible to minor, as the campaign was purely commercial and did not target critical systems or state functions.
The AI system (GPT-4o-mini) is used to perform a specific, well-defined function of generating a single customized marketing message based on direct inputs scraped from target websites. It operates on direct input without independent planning or context retention, acting as a tool within the larger automated Python framework.
Limited Risk: The AI system is used to generate synthetic text content (marketing messages) sent to users, which falls under the transparency obligations for AI-generated content where users should be informed they are interacting with or receiving AI-generated outputs.
The AI system was prompted to act as an assistant that generates marketing messages, specifically tailoring outreach text based on scraped website content.
The reports from SentinelOne and other cybersecurity outlets provide highly detailed technical breakdowns of the [name withheld]Bot framework, including its source code, API prompts, target metrics, and the specific AI model used (GPT-4o-mini). The facts are consistent across all sources, and OpenAI's response confirming the abuse and disabling of the API keys provides strong validation.
None. The evidence clearly points to a coordinated, financially motivated spam campaign leveraging LLMs to bypass filters.
Technology
Abuse of GPT-4o-mini API
Threat actors abused OpenAI API to generate unique spam at scale.
Hardcoded API Keys
Hardcoded keys allowed unauthorized, continuous access to LLM services.
Unique Content Generation
LLM-generated unique text bypassed traditional static spam filters.
Data Inputs
Scraped Website Content
BeautifulSoup scraped site text to feed custom context to the LLM.
Generic Spam Templates
Templates structured the LLM prompts to output tailored marketing spam.
Human Factors
Malicious Operator Intent
Threat actors intentionally bypassed protections to distribute spam.
Process and Methods
Automated Prompt Engineering
Prompt instructed the LLM to act as a helpful marketing assistant.
Concurrent Thread Execution
GUI allowed concurrent, automated targeting of thousands of websites.
Management
Lack of Proactive Key Detection
OpenAI relied on third-party security reports to identify API abuse.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
