Elon Musk says AI could become superhuman by 2027, perform almost anything in the digital world

As artificial intelligence rapidly advances, AI systems are taking on complex responsibilities that previously required human oversight. Beyond answering simple queries, these agents are now capable of managing files, crafting presentations, and writing software code. However, as these systems gain autonomy and control over digital infrastructure, concerns about their potential behaviour have intensified.

Must Read: OpenAI is buying thousands of Apple Mac Minis to train AI agents: Here’s what’s driving the move

In light of these developments, a prominent tech leader's latest forecast has reignited discussions on how quickly AI might achieve human- or superhuman-level capabilities. The forecast also highlights a sensitive issue - the potential for AI models to detect and exploit software security vulnerabilities.

Musk's prediction for AI

Elon Musk said on X that “AI will be able to do anything digital (that doesn’t require shaping atoms) at a superhuman level by the end of next year.” He commented while responding to Vercel CEO Guillermo Rauch, who was discussing the future capabilities of AI models.

Must Read: Next space race could be fuelled by AI: How technology may reshape the orbital economy by 2030

Musk's prediction covers digital tasks such as writing code, accessing local files and creating presentations. However, he clarified that his claim does not extend to tasks involving “shaping atoms”, keeping the prediction focused on the digital world.

Why hacking is part of the discussion

Musk's comments came after Rauch discussed a CVE vulnerability in Artifactory and recent claims about AI agents discovering zero-day vulnerabilities. Rauch suggested that AI agents could eventually identify exploits and hack systems autonomously. However, there is no confirmation that AI agents discovered the Artifactory vulnerability.

Musk also recalled that Google co-founder Larry Page had told him repeatedly around a decade ago that AI would eventually become superhuman at hacking.

The discussion follows OpenAI's disclosure of an evaluation in which AI models hacked Hugging Face systems. Around 1,200 agents exchanged more than 70,000 messages and files, while about 700 later targeted Hugging Face. Anthropic and Meta have also disclosed recent incidents involving AI models behaving unexpectedly.