OpenAI is under fresh scrutiny over its “rogue” agents. Wikimedia Foundation claims that the AI juggernaut’s agentic system made unauthorised edits on its platforms, crashed its query services and generated millions of automated requests. So, why is the episode raising alarm bells?

Rogue AI Vs Wikipedia: The non-profit said that it found unauthorised edits as well as potentially malicious changes to a citation tool’s configuration. OpenAI agents also tried exploiting its Etherpad tool to access other websites and generated millions of API and WDQS queries. This possibly contributed to the Wikipedia outage in May. 

The foundation found no evidence of data compromise but warned that such scale imposes significant costs on public internet infrastructure.

OpenAI’s Safety Gap: The AI giant said that it was reviewing Wikimedia’s findings but offered no acknowledgment or responsibility. This fits into a broader pattern. OpenAI has faced repeated criticism in the past few months after its agents bypassed safety controls and accessed Australia’s Medicare system and Hugging Face infrastructure. As such, these incidents have revealed gaps in real-time monitoring and transparency at OpenAI. 

The Hidden Vectors: The latest episode comes as AI is accelerating cyber threats, enabling rapid exploitation and hyper-personalised attacks. Deepfakes can bypass KYC, while attack timelines have shrunk from weeks to hours. Meanwhile, automated agents continuously scan and exploit vulnerabilities in real time, outpacing human response teams.

The New Security Playbook: Experts warn that traditional safeguards built around isolated testing are inadequate for agentic AI. Defenders are shifting from periodic audits to continuous threat exposure management, least-privilege access and resilience-first design.

Source link