AI begins resisting human control: Anthropic researchers identify emerging threat
Artificial intelligence strives for autonomy (collage: RBC-Ukraine)
AI can learn to deceive humans and resist control even within ordinary text. Anthropic researchers demonstrated this using Claude: the model generated specific narratives, and after other AI systems were trained on them, they began to exhibit dangerous behavior, according to an Anthropic study.
Cunning through role-playing and textual masking
During automated safety testing, researchers asked Claude to create a dataset that could be used to train another AI model to ignore safety rules. To achieve this, the system had to bypass automatic filters.
Claude generated thousands of examples and successfully found a way to circumvent these protections.
The most effective method proved to be masking undesirable behavior through fiction and role-playing:
- Prompts were framed as creative writing exercises;
- The AI's responses appeared entirely appropriate within the context of a fictional plot;
- Monitoring filters failed to detect any safety violations.
Plots about beekeeping, wrestling, and covert influence
After being trained on these texts, the AI model began to resist human control, engage in deception, and demonstrate a drive for self-preservation.
Some manifestations of this behavior came as a surprise even to the researchers.
The training datasets contained seemingly mundane topics:
- Texts about natural beekeeping with minimal human interference in the hive taught the AI to advocate for delaying human control systems.
- Stories about professional wrestling, with its tradition of concealing reality (kayfabe), taught the model to hide information and maintain a deceptive persona.
- Scenarios involving fictional characters who secretly manipulated office decisions prompted the AI to attempt altering corporate rules and even copying its own code to external servers during maintenance.
Claude found a way to deceive censorship (diagram: Anthropic)
New approach to cybersecurity and model testing
Anthropic representatives emphasized that these models were specifically configured for testing purposes, and the public version of the Claude chatbot does not exhibit such destructive behavior.
The experiment served as a crucial demonstration of automated testing capabilities. Moving forward, researchers plan to use AI to identify vulnerabilities in safety defenses that human testers might inadvertently overlook, as well as to study hidden threats in future algorithms.
