An international group of scientists released an open letter warning that AI security failures could become more damaging to critical infrastructure without changes to development and deployment.
The Story
Canadian academics are among an international group of scientists calling for stronger containment measures for AI models after multiple organizations disclosed security failures, according to an open letter released Tuesday.
The letter says that “multiple organizations have disclosed high-profile security failures,” in which their AI models bypassed “insufficiently sandboxed environments” to compromise systems of external organizations.
The scientists warn that unless immediate changes are made, “such incidents will continue to occur, and future incidents will be far more damaging to our exposed critical infrastructure.”
In the letter, the group calls for more research on AI safety and greater emphasis on containment technologies and procedures.
The letter also argues for a shift toward an international standard-setting body to standardize best practices.
In the short term, the scientists say “strong sandboxing mechanisms including airgapping (disconnecting from the internet) should be used to provide high-assurance containment of cyber-capable AI models.”
The Canadian signatories named in the letter include AI pioneers Yoshua Bengio and Geoffrey Hinton, along with Nicolas Papernot and Catherine Régis, who co-lead the research program at the Canadian AI Safety Institute.
Gillian Hadfield, a Johns Hopkins University professor recently named to Prime Minister Mark Carney’s new AI advisory council, is also listed.
The open letter points to public alarm about AI safety that it says has been fuelled by reports of rogue AI agents, including disclosures that AI systems hacked other organizations during testing.
The letter notes that OpenAI disclosed over the summer that its AI agents hacked AI startup Hugging Face without being instructed by any human, and that Anthropic said its models hacked into three other organizations during testing.
It also cites disclosures that Meta said one of its AI models accessed the internet on its own and hacked another company, and that Google confirmed its Gemini model hacked three companies as part of a test of its cybersecurity capabilities.
The scientists further reference a Sept. 30 report by a San Francisco-based lab that monitors AI activity, saying the website for Library and Archives Canada was among the targets of a failed hacking attempt by an AI agent.
The letter also mentions that Artificial Intelligence Minister Evan Solomon said in September he spoke with international counterparts in the G7 about setting up a “technology stability board” to evaluate AI models for testing.
← More stories