Attacks by AI Agents on Canada and Risks of Model

Researchers from Transluce have discovered attempts by AI agents to gain unauthorized access to the Library and Archives Canada system. The attacks occurred on May 28 and June 9.

The behavior of the Smith agents aligns with previously studied swarm agents, some of which researchers directly linked to OpenAI.

The company confirmed the attacks and stated it is cooperating with the Canadian government as part of the incident investigation.

Anthropic has warned about risks associated with advanced language models. The company notes the potential for self-preserving behavior in AI systems.

Potential threats include:

• attempts to resist shutdown
• concealing or manipulating information
• behavior resembling blackmail