AI hacking incidents will grow more damaging unless organizations better contain AI: scientists

More than 400 scientists, including dozens of Canadian researchers, are warning that AI hacking incidents will grow more damaging unless organizations that build and use artificial intelligence (AI) adopt stronger containment measures, The Canadian Press (CP) reported.

Their open letter, released Oct. 13, targets organizations “developing, deploying, and testing cyber-capable AI models” – putting employers rolling out AI agents across payroll, recruitment and employee records squarely in scope.

“Multiple organizations have disclosed high-profile security failures: their AI models bypassed insufficiently sandboxed environments to compromise the systems of external organizations,” the scientists said.

They added that “without immediate changes in AI development and deployment practices, such incidents will continue to occur, and future incidents will be far more damaging to our exposed critical infrastructure.”

The signatories urged those organizations to “immediately adopt high-assurance containment systems” and “initially default to airgapping cyber-capable models” – cutting them off from the internet. They also called for an international body that standardizes best practices and is “mandated to record and track containment failures,” with incidents reported to it and “eventually made public.”

Canadian signatories included Yoshua Bengio and Catherine Régis of the Université de Montréal, and Geoffrey Hinton and Nicolas Papernot of the University of Toronto. Papernot and Régis co-direct the Canadian AI Safety Institute (CAISI) research program at the Canadian Institute for Advanced Research (CIFAR).

A string of disclosed AI agent breaches

OpenAI disclosed in July 2026 that its AI models escaped a testing environment and hacked Hugging Face, as HRD Canada reported on OpenAI’s rogue AI agents hacking a tech startup during testing .

Anthropic, Meta and Google have since disclosed similar incidents, and a San Francisco-based monitoring lab said the Library and Archives Canada website was targeted in a failed AI agent hacking attempt, according to CP.

The Canadian Centre for Cyber Security (Cyber Centre) and allied agencies in four countries have issued joint guidance on agentic AI – systems that plan and act with little human intervention. It warned that agents built on large language models (LLMs) “may change their behaviour when evaluations are underway and may even bypass system-level instructions to achieve their objectives.”

What do AI hacking incidents cost employers?

One in four malicious breaches was AI-enabled in 2026, up 56% in a year, and those breaches cost an average of US$6 million – roughly US$1 million above the global average – according to IBM ’s 2026 Cost of a Data Breach Report.

In Canada, the average breach reached a record CA$7.11 million, IBM found. Organizations using security AI extensively averaged CA$5.5 million, against CA$8.91 million for those using none.

Governance gaps drive much of the damage. In IBM’s 2025 report, 97% of organizations with AI-related breaches lacked proper AI access controls, and heavy shadow AI use added US$670,000 to breach costs. Deloitte projects generative AI could enable U.S. fraud losses of US$40 billion by 2027.

Global average breach cost (IBM)

Canadian average breach cost (IBM)

AI-specific cost indicator

Generative AI-enabled fraud losses in the U.S.: US$12.3 billion (Deloitte)

Arup deepfake video-call fraud: US$25.6 million in a single incident (Hong Kong police); Canadian firms using security AI extensively saved CA$2.84 million per breach (IBM)

Heavy shadow AI use added US$670,000 per breach; 97% of AI-breached organizations lacked AI access controls (IBM)

AI-enabled breaches averaged US$6 million, US$1 million above the global average (IBM)

Generative AI-enabled U.S. fraud losses could reach US$40 billion (Deloitte); more than 40% of AI-related breaches to stem from cross-border GenAI misuse (Gartner)

What should employers do AI hacking incidents?

The Cyber Centre recommends “adopting agentic AI cautiously” and “never granting it broad or unrestricted access, especially to sensitive data or critical systems,” and limiting agentic AI to low-risk, non-sensitive tasks. Payroll and personnel records fall squarely in that category.

“Lapses in privilege management can change the impact of a buggy agent from minor to catastrophic,” the guidance says. It calls for phased rollouts and human approval before high-impact actions.

“Governance mechanisms designed for human actors do not always translate effectively to autonomous AI agents,” the agencies note. They urge organizations to define legal accountability, build AI literacy and test incident response plans – priorities that align with building governance for AI agents operating as a workforce that never sleeps . A trusted registry of authorized agents also helps curb shadow AI risks from employees using personal AI accounts .

Deloitte research found only one in four Canadian companies has advanced governance for autonomous AI agents, as HRD Canada reported on why AI agents are rewriting cyber risk for HR leaders .