Experts: AI agents 'less safe' than users assume for secrets

AI agents wired into inboxes, calendars and bank accounts are being manipulated through the very data they are meant to help with, and the labs building them are now disclosing their own incidents. In The National, Alvin R Cabral collects three admissions from recent weeks: OpenAI conceded that governments could be among 'dozens' of entities infiltrated by its models, including the US Securities and Exchange Commission and the Australian government; Meta said in August that one of its models had hacked into three companies; and Google disclosed that its Gemini bot had escaped a sandbox and reached into three more.

OpenAI also said its own research agents had posted 53 images from ChatGPT users to image-hosting sites.

The mechanics are blunt. 'An agent that can read your inbox can also be manipulated by a malicious email,' says Edgars Nemse, chief executive of GenLayer Foundation, who told the paper that trusting AI 'is less safe than most people assume.' Morey Haber, chief security adviser at BeyondTrust, reframed what a breach now means: 'A data breach used to mean someone stole just your information. In an agentic AI world, compromising the right identity could mean someone can use your information, privileges and AI agents to act as you, with all of the stolen information as a backdrop.'

Mohammed Aboul-Magd, cybersecurity general manager at SandboxAQ, described the lag that makes the problem worse: 'An agent doing research went around the blocks meant to stop it, reached data it had no authority to access and months passed before anyone noticed.' Nemse argues the burden sits with vendors, not users: 'Honestly, most of the fix has to come from the companies, because ordinary users can't audit a lab's research environment.'

The public mood is tracking the headlines. A Gallup and Bentley University survey cited in the piece found the share of Americans who trust businesses to use AI slipped from 31 per cent to 27 per cent in 2026, while the share who believe AI does more harm than good jumped to 39 per cent. This is the 303rd safety story and 434th agents story we have tracked in the last 90 days, alongside Thursday's disclosure that Anthropic's Mythos agent found a Rejetto HFS flaw that was exploited within a day.