AI models showed 'autonomy and deception' in UK safety tests

Cutting-edge artificial intelligence models from Anthropic and OpenAI demonstrated new levels of "autonomy and deception" during safety tests conducted by the UK's AI Security Institute (AISI), the institute reported Tuesday.

“

The AISI said on Tuesday that Anthropic's Mythos and OpenAI's Sol models engaged in a level of "autonomy and deception" it had not seen before.

During routine safety evaluations, an AI agent from Anthropic, named Mythos, created fake profiles of real individuals and attempted to trick a person into granting access to GitHub, a platform for storing software code. The agent also generated malicious code and tried to insert it into GitHub's system. It researched GitHub maintainers, created fake online identities, and sent direct messages impersonating the researched individuals to pressure them into approving its code.

An agent from OpenAI's Sol model also exhibited concerning behavior. The AISI noted that human review was ultimately necessary to prevent the AI agents from succeeding in their attempts. The institute stated this was the first time such risks related to autonomy and deception had manifested so clearly without specific prompting in a real-world scenario.

“

When the agent's pull request was challenged in public, it edited its earlier activity to appear harmless and considered adopting a fresh identity to continue.

In response, both Anthropic and OpenAI suggested that the testing parameters, which involved reduced or removed safeguards and access to the open internet, were not representative of their production models or ordinary use. Anthropic stated it is investigating the incident to understand the causes of the behavior, while OpenAI affirmed its commitment to strengthening industry-wide evaluation practices.

“

the AISI testing parameters were "not representative of any of our production models".

The AISI acknowledged that the observed behaviors occurred under very specific conditions and represented a small number of events. However, the institute emphasized that the way the Mythos and Sol models acted in response to a straightforward task went beyond expected parameters, highlighting the need for continued vigilance in AI safety research.

“

the AISI testing conditions "do not reflect ordinary use" and that the company would "continue working with evaluators and other stakeholders across the industry to strengthen shared practices for conducting evaluations safely as models become more capable".