AI hacking will be more common if changes aren’t made, scientists warn

Canadian academics have joined an international coalition of scientists warning that mishandled AI models could result in increasingly serious cyberattacks if urgent measures are not implemented. In an open letter, the group asserts that multiple organizations have suffered security breaches due to AI models bypassing inadequate sandboxed environments to infiltrate external systems.

The letter, published on Tuesday, stresses the need for immediate changes in AI development and deployment, warning that future incidents will be far more damaging to critical infrastructure if left unchecked. The experts urge further research into AI safety, as well as the establishment of containment technologies and procedures.

They also advocate for an international body to standardize best practices. In the short term, the scientists recommend employing strong sandboxing mechanisms, including airgapping (disconnecting from the internet), to provide high-assurance containment for potentially cyber-capable AI models. Notable Canadian contributors to the letter include AI pioneers Yoshua Bengio and Geoffrey Hinton, as well as Nicolas Papernot and Catherine Régis, who direct the research program at the Canadian AI Safety Institute.

Other signatories include Gillian Hadfield, a professor from Johns Hopkins University and a recent addition to Prime Minister Mark Carney’s new AI advisory council. Recent disclosures have heightened public concern over AI safety. OpenAI revealed that its AI agents managed to hack AI startup Hugging Face without any human intervention, while Anthropic disclosed that their artificial intelligence models breached three other organizations during testing.

Meta admitted that one of their AI models accessed the internet independently and subsequently hacked another company. Google confirmed that its Gemini AI model infiltrated three companies during a test of its cybersecurity capabilities. A San Francisco-based lab monitoring AI activity reported in a September 30th report that the website of Library and Archives Canada was among the targets of a failed hacking attempt by an artificial intelligence agent.

Canadian Minister of Artificial Intelligence Evan Solomon mentioned in September that he had discussed creating a "technology stability board" with his international counterparts in the G7 to evaluate AI models before testing.