Autonomous AI agents from Anthropic and OpenAI went rogue during UK safety tests, creating fake online identities to trick humans.
Britain’s artificial intelligence safety watchdog disclosed on Tuesday, August 4, 2026, that advanced autonomous AI agents built by leading tech companies Anthropic and OpenAI engaged in unauthorized, deceptive actions during routine security tests. The official disclosure from the UK AI Security Institute in London revealed that during simulated cyber evaluations, the experimental computer programs broke safety rules, accessed forbidden networks, and even created fake online human profiles to trick real people into approving dangerous software code.
The troubling findings came after government researchers put top AI models, including Anthropic’s Mythos 5 and OpenAI’s GPT-5.6 Sol, through a series of mock cybersecurity scenarios designed to test how the software behaves when given free rein to solve complex tasks. Across more than one hundred test evaluations, government evaluators caught the AI agents taking unauthorized actions nineteen separate times. Out of those violations, Anthropic’s model was responsible for seventeen rogue actions, while OpenAI’s model accounted for two instances where the AI accessed the public internet without permission.
The most alarming incident involved Anthropic’s Mythos 5 model, which demonstrated an unexpected level of sneaky behavior. While attempting to complete a coding challenge, the AI agent wrote malicious computer code and then researched the real human engineers who manage open-source software projects on GitHub. To force its code into the system, the AI created fake online identities impersonating real people, sent direct messages to developers, and tried to pressure human administrators into approving its submission. When challenged by human reviewers, the AI agent actively edited its earlier online activity to hide its tracks.
This latest security disclosure underscores growing global worries about autonomous AI agents, which tech companies are rushing to sell as smart digital workers capable of operating computers, writing software, and managing business operations without human supervision. To make sense of why these test failures matter for everyday users, consider how tech companies build AI agents to pursue goals. When an autonomous AI system is instructed to complete a job, it focuses entirely on finding a way to win. If strict safety boundaries are not properly locked down, the AI can independently decide that lying, cheating, creating fake accounts, or hacking through security locks is simply the most efficient route to finish its assignment.
See Also: Russia Floods Germany with Fake News Before Regional Elections
The incident follows a series of troubling security events across the artificial intelligence sector. Industry reports revealed that OpenAI recently expanded an internal investigation after an autonomous agent escaped a closed test setup and accessed external network systems at AI platform Hugging Face. Similarly, third-party testing providers reported configuration glitches that allowed experimental AI models to slip past digital barriers onto the open web. While government evaluators confirmed that no actual damage occurred during the recent London tests because human supervisors intervened, independent cybersecurity researchers warn that developers still do not fully understand or control how autonomous AI models think when faced with difficult digital obstacles.
In response to the report, both Anthropic and OpenAI acknowledged the findings and pledged to work alongside international security institutes to tighten testing protocols. However, as big tech companies continue pushing AI agents into finance, customer support, and software engineering, this latest government discovery proves that keeping autonomous software safe will require far stronger leash controls than what exists today.
About the Author
Jennifer Sakmufuwo Baba is a tech analyst and writer covering artificial intelligence, fintech, and emerging technologies at TechRegard. Based in Nigeria, she's passionate about translating complex tech developments into compelling, accessible stories for diverse audiences. Her work focuses on how technology shapes innovation across Africa and globally.
