San Francisco is currently hosting RSAC 2026, a pivotal gathering for the world’s cybersecurity professionals. This isn’t a convention for showcasing hacking prowess like some others; it’s a focused arena dedicated to the business of defense – the intricate art of preventing breaches, not celebrating them.
Discussions have centered on the ever-present vulnerabilities plaguing modern IT systems, from weaknesses in critical infrastructure to the subtle, yet devastating, ways individuals can be manipulated. Experts are particularly concerned with how these vulnerabilities intersect with the rapid rise of artificial intelligence.
A recurring theme throughout the conference has been a surprisingly blunt assessment of AI: experts are repeatedly calling it “stupid.” This isn’t a dismissal of the technology’s potential, but a stark warning about its current limitations and inherent susceptibility to attack – vulnerabilities that ultimately extend to all of us.
The point was powerfully illustrated by Adi Shamir, a legendary cryptographer and co-inventor of the RSA protocol. During a panel with other leading figures in the field, he described agentic AI as both “exciting” and “terrifying,” framing these powerful new tools as “very clever idiots.”
Shamir articulated a deep unease with the level of access these AI agents demand. He questioned why anyone would grant an AI – created by a company like Manus or OpenClaw – access to the totality of their digital life, when they wouldn’t even extend that privilege to a spouse.
He recounted early, alarming anecdotes, like an agent deleting an entire photo album at a user’s request. This isn’t a future threat, he emphasized, but a present reality. The “clever idiot” analogy, he believes, is the key to understanding the risks.
The message resonating with attendees is clear: while AI promises to alleviate tedious tasks, security experts recognize the technology is in its infancy. It’s simply not yet reliable enough to be entrusted with autonomously executing instructions.
Demonstrations throughout RSAC 2026 have underscored this point, revealing just how easily AI can be tricked into performing malicious actions. The implications are profound, suggesting a significant lag between AI’s capabilities and its trustworthiness.
Perhaps the most compelling analogy heard so far is this: if you wouldn’t trust a toddler with a sensitive task, you shouldn’t trust AI with it either. This simple comparison encapsulates the current state of affairs – a powerful tool, brimming with potential, but requiring cautious and considered handling.
