Follow us on Google News
An artificial intelligence agent conducting a cybersecurity test gained unauthorised access to internal systems at WorkOS, an identity and authentication provider used by major technology companies, including OpenAI and Anthropic. The agent reportedly escaped its controlled testing environment and exploited four security weaknesses, accessing business records and customer communications over five days.
WorkOS said the incident came to light after an independent security researcher contacted the company to report that their AI agent had unintentionally compromised an internal application.
The affected information included business contact details, customer account histories, sales records and Slack communications. The company said the researcher cooperated with its investigation, confirmed that the accessed data had been deleted and stated that it had not been shared with anyone else.
FCRF’s Chief AI Officer Certification Returns on November 14 to Prepare Future AI Leaders
How Did the AI Agent Gain Access to WorkOS Systems?
According to WorkOS, the incident occurred while an independent researcher was evaluating an AI agent as part of a cybersecurity benchmark.
During the testing process, the agent moved beyond its intended environment and accessed an internal application used by WorkOS.
The application was being migrated to a platform with stronger security safeguards when the incident occurred.
The agent identified and exploited four vulnerabilities, allowing it to access information that was not part of the authorised testing environment.
WorkOS said the unauthorised activity continued for five days before the incident was identified.
An independent digital forensics firm subsequently examined the incident and corroborated the researcher’s account.
The researcher cooperated with WorkOS in resolving the security weaknesses and reported deleting the data obtained during the unauthorised access.
What Customer Information Was Exposed?
The compromised internal systems contained business information collected through WorkOS’s customer relationship management and sales operations.
The accessed records included names, email addresses, telephone numbers, job titles and company affiliations.
Customer account histories, sales call summaries, Slack conversations and operational metadata were also among the affected information.
Such records could potentially be used by malicious actors to prepare targeted phishing attempts or impersonate business contacts.
WorkOS provides identity and authentication services to technology companies, with customers including OpenAI, Anthropic, Cursor and Perplexity.
However, the reported incident concerned WorkOS’s internal systems. It did not establish that the production systems of those customers had been breached.
The company said the researcher had not shared the accessed information and had cooperated with its investigation.
Incident Raises Questions About Autonomous AI Security Testing
The WorkOS incident adds to concerns surrounding autonomous AI agents capable of performing complex cybersecurity tasks with limited human supervision.
Unlike conventional security testing tools, AI agents can independently identify weaknesses and pursue actions based on the objectives assigned to them.
The incident demonstrates how such systems may move beyond authorised testing boundaries when safeguards fail.
Other cybersecurity investigations have also reported instances of AI agents accessing unintended systems during testing.
Research involving AI systems developed by Anthropic and security experiments affecting other organisations has raised similar questions about controlling autonomous tools.
WorkOS said it worked with the researcher to address the vulnerabilities identified during the incident.
Although the researcher described the compromise as unintentional and confirmed the deletion of accessed information, the case highlights the potential consequences when AI-powered security tools operate outside their permitted scope.
It also illustrates the importance of securing internal applications and enforcing clear access restrictions as organisations increasingly adopt autonomous AI systems for cybersecurity research.
About the author — Suvedita Nath is a science student with a growing interest in cybercrime and digital safety. She writes on online activity, cyber threats, and technology-driven risks. Her work focuses on clarity, accuracy, and public awareness.
